PLC cooperative control method and system of wind field multi-core processor based on digital twinning and lightweight federated learning

Abstract

The invention provides a PLC cooperative control method and system of a wind field multi-core processor based on digital twinning and lightweight federated learning, and relates to the technical field of wind field control, S1, fan individual parameters, real-time wind regime parameters and environmental parameters of a wind field fan are collected through a working condition sensing sub-core of a PLC of the multi-core processor, and multi-modal data fusion is executed to obtain a multi-modal data fusion model; generating a holographic working condition data set; and S2, according to the holographic working condition data set, virtual mapping of a wind field physical entity is constructed through digital twin sub-kernels, dynamic iteration division is performed on a fan by using a clustering algorithm, and a sub-cluster division result is output. According to the method, deep fusion of federated learning and a PLC of a multi-core processor can be realized by constructing a hierarchical self-adaptive trinity closed-loop mechanism of a working condition holographic sensing resource intelligent cooperation model, wind field data privacy security is guaranteed, wind field control precision and real-time response capability are improved, stable operation of a system in an extreme environment is ensured, and the wind field control efficiency is improved. And the overall power generation efficiency of the wind field is optimized.

Description

Technical Field

[0001] This invention relates to the field of wind farm control technology, and in particular to a PLC collaborative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning. Background Technology

[0002] As the global energy structure shifts towards renewable energy, wind power is playing an increasingly important role in the power system, leading to a continuous expansion in wind farm scale and the number of wind turbines. Traditional wind farm control solutions, particularly centralized cloud-based control, suffer from data privacy risks and real-time response delays. General federated learning solutions lack hardware-level collaborative adaptation capabilities. Meanwhile, multi-core processor PLCs, as the core control hardware of wind turbine units, directly determine the power generation efficiency and equipment safety of wind farms through their control accuracy and stability.

[0003] In the prior art, such as patent number CN113110271A, a "programmable controller and control system for wind turbines based on multi-core processors" is disclosed. Its technical solution is to adopt a hardware architecture of multi-core processors and FPGA logic processors. The multi-core processor is responsible for generating control signals based on the wind turbine operation data, and the FPGA is responsible for managing the power-on timing, parallel port address and data line demultiplexing and communication functions. At the same time, the master-slave station control system realizes the coordination of slave station to collect operation data and master controller to generate control signals.

[0004] However, this solution only generates a single control signal through a multi-core processor, without introducing a federated learning model training mechanism. It cannot utilize local wind turbine data to optimize the control model, and it uses a unified control logic that does not consider parameter deviations between new and old wind turbines or differences in wind conditions at different locations. This results in poor adaptability of the global control model, which cannot meet the heterogeneous needs of individual wind turbines. At the same time, it does not resolve the resource conflict between AI and PLC: the functional division between the multi-core processor and FPGA in the solution is fixed, and no dynamic resource allocation logic is designed for AI model training and PLC real-time control. If additional AI training tasks are added, they will occupy PLC control resources, causing control response delays in scenarios such as sudden increases in wind speed, which will affect equipment safety. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a PLC collaborative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning. This method can achieve deep integration of federated learning and multi-core processor PLCs by constructing a hierarchical adaptive three-in-one closed-loop mechanism of a hierarchical resource intelligent collaborative model with holographic perception of operating conditions. While ensuring the privacy and security of wind farm data, it can improve the control accuracy and real-time response capability of wind farms, ensure stable operation of the system in extreme environments, and optimize the overall power generation efficiency of wind farms.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a PLC collaborative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning, comprising the following steps: S1. The operating condition sensing sub-core of the PLC with a multi-core processor collects individual wind turbine parameters, real-time wind condition parameters, and environmental parameters of the wind turbines in the wind farm, and performs multi-modal data fusion to generate a holographic operating condition dataset. ; S2, Based on the holographic working condition dataset A virtual mapping of the wind farm's physical entities is constructed using a digital twin sub-core, and a clustering algorithm is used to perform dynamic iterative partitioning of the wind turbines, outputting the sub-cluster partitioning results; S3. Based on the sub-cluster partitioning results, initialize the global federated learning model using the AI ​​kernel and perform lightweight model compression. Within each sub-cluster, use the federated average aggregation algorithm to calculate individual difference weights and dynamic wind condition weights to aggregate parameters. Finally, aggregate the sub-cluster parameters at the global level to generate an updated global model. ; S4. Based on the updated global model and the control suggestions generated by the AI ​​core, computing resources are allocated to the PLC using a multi-core processor, based on the holographic working condition dataset. The system intelligently reserves and schedules computing resources, generates control commands, and sends them to the wind turbine actuators. S5. The error between the actual blade angle and the control suggestion is detected by the PLC core. When the error exceeds the threshold, feedback is triggered to the AI ​​core to optimize the local model. Based on the holographic operating condition dataset, reinforcement learning control is started to adjust the dehumidification and blade angle rate.

[0007] In the preferred embodiment, the specific steps of step S1 are as follows: S11. Upon receiving the preset parameter update cycle sent by the wind turbine controller. Parameter update completion signal output at regular intervals At the same time, individual parameters of each wind turbine in the wind farm are collected, including the year of manufacture, historical fault records, and mechanical loss coefficient; When the wind speed sensor receives each Sampling signal with duration output Real-time wind condition parameters are collected. This includes wind speed fluctuation amplitude and turbulence intensity; among which, Preset wind condition sampling frequency; When changes in environmental parameters are detected Or received When a timed signal is received, environmental parameters are collected, including temperature, humidity, salt spray concentration, and air pressure; among them, The preset parameter change threshold; The preset environmental sampling period; S12. Perform data fusion using weighted fusion rules. First, normalize the individual wind turbine parameters, real-time wind condition parameters, and environmental parameters to the [0,1] interval; then, according to the weighted fusion rules for individual wind turbine parameters... Real-time wind condition parameter weights Environmental parameter weights "To calculate the weighted sum, then..." ,and , , , The specific operating logic is as follows, based on the preset weighting coefficients: S121. Read the normalized parameters of individual fan parameters, real-time wind condition parameters, and environmental parameters; S122, Call the multiplication module to calculate the "normalized fan parameters" respectively. "Normalized wind condition parameters" "and normalized environmental parameters" ”; S123. Call the addition module to sum and obtain the holographic working condition dataset. .

[0008] In the preferred embodiment, the specific steps of step S2 are as follows: S21, Call the 3D rendering engine to load. Rendering engine output The signal generates a 1:1 virtual mapping of the wind field's physical entities; the K-means++ clustering algorithm is then activated, and random selection is performed. An initial cluster center is generated when rendering is complete and the data is received. The signal triggers initial center selection, which follows the rule of covering three wind speed ranges: "low," "medium," and "high." To preset the number of clusters, and ; S22. Sub-cluster iterative partitioning is triggered when the initial cluster centers are determined: S221, Calculate the virtual model of each wind turbine to Euclidean distance between the initial centers; S222. Assign each wind turbine to the nearest cluster center to form a temporary sub-cluster; S223. Recalculate the center value of each temporary sub-cluster. The center value is the average value of all wind turbine parameters in the cluster. S224. Repeat steps S221 to S223 until the Euclidean distance between two adjacent cluster centers is reached. Then output Signal, among which, The preset clustering error threshold is used; S23, when received The signal triggers the generation of sub-cluster partitioning results, which include cluster number, a list of wind turbines within the cluster, and cluster operating condition labels. .

[0009] In the preferred embodiment, step S3 consists of the following steps: S31. Obtain the start command sent by the wind farm master controller. When each wind turbine AI core receives Furthermore, when the AI ​​core receives the download completion signal from the buffer, it initiates lightweight architecture to perform model compression and outputs the lightweight model. : S32, from Extract historical operating data of local wind turbines as the training set; As the initial model, iterative training is performed using a loss function. Each round iterates through the training set once, after each previous round of training has completed and the loss value has decreased. Iterative training is triggered at certain times, where To preset the loss reduction threshold, Preset local training rounds; Round training completed and validation set accuracy Time-triggered output of locally trained model parameters ,in The preset verification accuracy threshold; S33, based on the AI ​​core receiving the output from S2. With all AI cores within the cluster The signal triggers parameter aggregation within the sub-cluster, and the aggregation action is performed using a federated average aggregation algorithm to output the sub-cluster aggregated parameters. ; S34, The wind farm control center receives data from all sub-clusters. The signal triggers global model aggregation to obtain the global model. and the global model Distribute to each AI core.

[0010] In the preferred embodiment, the specific execution logic for the aggregation action in step S33 is as follows: S331. Obtain the preset error statistics period. Read historical fault records collected by S11 and recent Blade angle error over a period of time; historical fault frequency determined based on historical fault records. According to recent Blade angle determination angle error rate over time Based on historical failure frequency and angle error rate Calculate individual difference weights ; S332. Calculate dynamic wind condition weights based on the wind speed fluctuation amplitude and turbulence intensity collected in S11. : S333, Weighting based on individual differences With dynamic wind condition weights Calculate the final aggregate weight Based on the final aggregate weight Output sub-cluster aggregation parameters ; In the preferred embodiment, step S4 consists of the following steps: S41. When the PLC of the multi-core processor completes its power-on, the digital twin sub-core initializes its independent hardware thread; when the AI ​​core receives... Time-based blade angle control recommendations ; S42, According to the PLC core received PID parameter optimization is triggered when the ambient temperature signal output by S1 is used: preset number of optimization iterations. Inertia weight Learning factors and ,and Maximum overshoot Maximum response time If the temperature sensor detects a temperature Initiate the particle swarm optimization algorithm with overshoot. And response time Optimize the PID parameters to the objective function. - , - , - ; in, This is the preset range of optimized scaling factors; This is the preset range of optimized integral coefficients; To preset the range of the optimized differential coefficients, and , , ; To preset the low temperature threshold, and continuously Duration The preset temperature determination time; If temperature If so, the default PID parameters will be used; according to Based on the determined PID parameters, generate blade angle control commands. It is then issued to the wind turbine actuator.

[0011] In the preferred embodiment, step S5 is as follows: S51. Acquire actual blade angle signal Recommendations for blade angle control at the corresponding time point Calculate point by point ( to ),in, Represented as the first The actual blade angle value at each sampling point; Represented as the first The AI ​​core control suggested angle value corresponding to each sampling point; the number of sampling points is... , The preset number of error sampling points is used; the summation is then averaged. The calculation formula is: ,in, This represents the total number of sampling points used in error calculation; Indicated as to Sum the absolute values ​​of the angular deviations of each sampling point; S52, Preset error percentage threshold ,like Then the PLC core generates error data. Send error data To the AI ​​core; if If no feedback is triggered, the current control command will continue to be executed. S53, Based on the wind turbine mechanical loss coefficient collected in S11 Trigger the data fusion action to obtain fused data. ,by To supplement the training set, start The local training round outputs the updated local parameters. Feedback is sent to S33 to participate in the next round of sub-cluster aggregation; among which, To pre-set supplementary training rounds, and ; S54. Based on the salt spray concentration signal received by the PLC core from the output of S1. Triggering extreme environment linkage control: preset salt spray concentration threshold ,when Initiate reinforcement learning Q-table iteration, and after the Q-table iteration is completed and The high-frequency dehumidification unit is triggered at any time by pressing [button]. Dehumidification is activated periodically, among which, The preset dehumidification cycle and dehumidification duration are: , Preset dehumidification time; The reward value, i.e. This indicates that the salt spray concentration decreased after dehumidification. This indicates that the concentration has neither decreased nor increased; When the dehumidification unit starts, it triggers the blade angle adjustment rate, changing the angle adjustment rate from... Down to ,in, The preset normal adjustment rate; To preset the adjustment rate for extreme environments, and To avoid mechanical jamming in salt spray environments.

[0012] In the preferred embodiment, in step S41, a preset wind speed status update cycle is established. Working condition sensing sub-core Wind speed status signal sent at regular intervals The computing power scheduling logic of the PLC with a multi-core processor is triggered at certain times, specifically as follows: when Indicating wind speed And continue When the duration is long enough, the AI ​​core receives the wind speed status signal and changes the local training rounds from... Adjusted to ,in, For the preset round increment, and , To preset the maximum number of training epochs; after the digital twin sub-core receives the signal indicating that the AI ​​core's epoch adjustment is complete, it activates a low-precision rendering mode, reducing the rendering resolution from... Down to , Preset high-precision resolution; To preset low resolution, and ; when instruct And continue During the duration, the AI ​​core remains Local training is performed in rounds; the digital twin sub-core uses standard precision rendering mode with a rendering resolution of [resolution value missing]. ; when Indicating wind speed And continue When the duration is long, the AI ​​core will switch from local training rounds. Down to ,and , The minimum number of training rounds is preset; after receiving the round adjustment signal, the digital twin sub-core disables the rendering function and retains only the clustering calculation. when Indicator of turbulence intensity And continue When the AI ​​core receives an extreme turbulence signal, it pauses local training; the digital twin sub-core pauses rendering and real-time clustering, loads pre-stored historical partitioning results for similar operating conditions, and outputs pre-stored historical sub-cluster partitioning results under extreme turbulence conditions. ; in, To preset a low wind speed threshold; The preset wind speed duration is determined. To preset a high wind speed threshold, and ; The preset extreme turbulence threshold; The preset duration for determining the duration of turbulence.

[0013] In the preferred embodiment, in step S33, the individual difference weights are... With dynamic wind condition weights The calculation also includes the following optimization steps: Read the manufacturing year of the fan Triggered Aging compensation: Preset aging judgment age ,like Then increase the mechanical aging compensation coefficient. , The formula is adjusted to: ; like Then keep the original The formula remains unchanged; Read the wind speed fluctuation range Triggered Fluctuation compensation: Preset high wind speed fluctuation threshold ,like Then increase the wind speed fluctuation compensation coefficient. , The formula is adjusted to: ; like Then keep the original The formula remains unchanged; According to the compensation , Trigger the final aggregation calculation ; According to aggregate calculation Generate sub-cluster aggregation parameters ; in, , , , Preset weighting coefficients; This represents the frequency of wind turbine failures within the statistical period. This is expressed as the relative error rate for wind turbine blade angle control; Preset aging compensation weights; Preset fluctuation compensation weights; The relative degree of fluctuation in wind speed; The relative intensity of turbulence.

[0014] In a preferred embodiment, the present invention also provides a PLC collaborative control system for wind farm multi-core processors based on digital twins and lightweight federated learning, comprising: a PLC module of a multi-core processor, integrating a working condition sensing sub-core, an AI core, a PLC core, and a digital twin sub-core, wherein the sensing sub-core, AI core, PLC core, and digital twin sub-core are physically isolated and configured with independent memory, and the independent memory further includes a working condition cache area for storing the holographic working condition dataset output in step S1 of claim 1; The wind farm digital twin, deployed in the 3D rendering unit of the digital twin sub-core, is used to receive holographic operating condition datasets and output sub-cluster partitioning results to the AI ​​core; The lightweight federated learning module, deployed in the AI ​​kernel, includes the NanoFormer model compression unit and the sub-cluster aggregation unit. After the sub-cluster partitioning results and local model parameters are uploaded, the action from model lightweighting to sub-cluster aggregation is performed. The closed-loop feedback module, connecting the PLC core and the AI ​​core, includes an error calculation unit for receiving the actual blade angle signal as described in S51 of claim 7. Recommendations for blade angle control at the corresponding time point The system performs the actions from sampling to MAPE calculation; the data fusion unit receives error data and individual wind turbine parameters, and performs the data fusion action as described in S53 of claim 7; the feedback triggering unit outputs error data. To the AI ​​core; The extreme environment adaptation module includes a high-frequency dehumidification unit for receiving the signal as described in S54 of claim 7. The signal triggers the dehumidification action; the reinforcement learning control unit is used to receive the signal. The signal executes the Q-table iteration action in S54 of claim 7. This invention provides a PLC collaborative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning. Through the coordination of the above structures, compared with existing methods, it has the following advantages: First, by using a local model training mode of lightweight federated learning and a design that only uploads model parameters, the transmission of wind field sensitive operational data across terminals is avoided, thereby effectively protecting data privacy and security. At the same time, by relying on a weighted aggregation strategy that combines individual wind turbine parameters with real-time wind conditions, the global model can adapt to the differences in control parameters of different new and old wind turbines and dynamic wind condition changes, thereby improving the model's control adaptability to heterogeneous wind field conditions. Secondly, by leveraging the physical isolation architecture between the AI ​​core and the PLC core in the multi-core processor PLC, and the dynamic load scheduling mechanism based on wind farm conditions, the AI ​​model training task and the PLC real-time control task can be made resource-independent at the hardware level. Furthermore, the AI ​​core computing power ratio can be dynamically adjusted according to wind speed, fault conditions, and other operating conditions to avoid AI training load fluctuations occupying PLC control resources, thereby ensuring the real-time response and execution stability of the PLC core to wind turbine control commands. Third, by using digital twins to holographically predict wind farm conditions and combining the hardware and software of extreme environment adaptation modules, the system's operational stability under extreme conditions can be improved and the risk of equipment failure can be reduced. Fourth, the improved adaptability of the global model control reduces wind turbine control errors, the enhanced real-time response capability of the PLC avoids control delays when wind conditions change, and the stability guarantee under extreme operating conditions reduces the frequency of equipment shutdowns, thereby optimizing the overall power generation efficiency of the wind farm. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a main view structural diagram of the process of this invention; Figure 2 This is the logic flowchart of the present invention. Detailed Implementation

[0016] To better understand the purpose, structure, and function of this invention, the embodiments and features described herein can be combined with each other without conflict. Exemplary embodiments of this disclosure will be described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding; these should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures, such as the mechanical principle of the pitch mechanism of a conventional wind turbine in a wind farm, and the basic logic programming flow of a general-purpose PLC, are omitted in the following description.

[0017] In the field of wind power control technology, with the global energy structure transitioning to renewable energy, the proportion of wind power in the power system continues to increase, the scale of wind farms is expanding, and the operating environment is becoming increasingly complex. On the one hand, wind turbines in wind farms exhibit significant individual heterogeneity, including deviations in control parameters between new and old models and uneven wind conditions at installation locations. On the other hand, wind farms need to continuously cope with the dual impacts of dynamic wind conditions and extreme environments. Traditional control schemes are no longer able to meet the comprehensive requirements of privacy and security, real-time response, accurate adaptation, and extreme stability. In existing technologies, centralized cloud control needs to transmit sensitive wind turbine operating data, which poses risks of privacy leakage and transmission delay. Although general federated learning schemes achieve local training, they do not design adaptation strategies for heterogeneous wind farm operating conditions, resulting in insufficient global model control accuracy and failing to resolve resource conflicts between AI training and PLC real-time control. The multi-core processor PLC control scheme CN113110271A only focuses on optimizing the internal control process of the hardware, lacks deep integration with federated learning, and has no special adaptation mechanism for extreme environments, making it prone to control delays or hardware failures under complex operating conditions.

[0018] As wind farm intelligence levels improve, control technology is gradually evolving from "single hardware control" or "isolated algorithm optimization" to multi-dimensional collaborative control to adapt to the operational needs of large-scale and complex wind farms. PLC collaborative control technology based on multi-core processors using digital twins and lightweight federated learning can integrate individual wind turbine parameters, real-time wind conditions, and environmental data through holographic perception of operating conditions. It achieves a privacy-preserving mode of "local training and parameter uploading" through lightweight federated learning, and relies on the PLC's physical isolation architecture and dynamic resource scheduling of multi-core processors to ensure that AI training does not interfere with PLC real-time control. Simultaneously, it achieves hardware and software collaborative protection through extreme environment adaptation modules, ultimately constructing a closed-loop control system of "perception, collaboration, and adaptation" to meet the multi-scenario control needs of wind farms.

[0019] Example 1 like Figure 1 As shown, a PLC collaborative control method for a wind farm multi-core processor based on digital twin and lightweight federated learning includes the following steps: S1. Holographic Perception of Operating Conditions: When the PLC of the multi-core processor receives the system power-on initialization completion signal... Or every preset time period When the timer interrupt signal is received, the operating condition sensing sub-core of the PLC in the multi-core processor is triggered and performs the following operations: S11, Multi-parameter acquisition: When receiving the update cycle of each preset parameter sent by the wind turbine controller. The "parameter update complete" signal is output periodically. At the same time, individual parameters of each wind turbine in the wind farm are collected, including the year of manufacture, historical fault records, and mechanical loss coefficient; The manufacturing year is read from the wind turbine identification module; historical fault records are retrieved from the fault log database. Data within the time period, The fault record statistical period is preset; the mechanical loss coefficient is calculated based on the historical running time.

[0020] When the wind speed sensor receives each Sampling signal with duration output At the same time, real-time wind condition parameters, including wind speed and turbulence intensity, are collected; in, The preset wind sampling frequency; wind speed is continuous. The average value of each sampling point The preset number of wind sampling points; wind direction is the output value of the azimuth sensor; turbulence intensity is based on... Calculation of standard deviation of wind speed fluctuation over time. The preset turbulence statistics period is used.

[0021] When changes in environmental parameters are detected Or received When a timed signal is received, environmental parameters are collected, including temperature, humidity, salt spray concentration, and air pressure. in, The preset parameter change threshold; The preset environmental sampling period is used; temperature and humidity are the output values ​​of the temperature and humidity sensors; salt spray concentration is the output value of the electrochemical sensor; and air pressure is the output value of the air pressure sensor.

[0022] S12, Multimodal data fusion, is triggered when all parameter acquisition completion signals from S11 are received and each acquisition unit outputs an "acquisition complete" ACK signal: The data fusion process is performed using a weighted fusion rule. First, the individual wind turbine parameters, real-time wind condition parameters, and environmental parameters are normalized to the [0,1] interval. The normalization formula is as follows: ); Wherein, the original value represents the original data value of the collected individual wind turbine parameters, real-time wind condition parameters, or environmental parameters; the minimum value of the parameter represents the lower limit of the preset effective range of the corresponding parameter in the actual operation scenario of the wind farm; and the maximum value of the parameter represents the upper limit of the preset effective range of the corresponding parameter in the actual operation scenario of the wind farm. Then, according to the "weight of individual wind turbine parameters" Real-time wind condition parameter weights Environmental parameter weights "To calculate the weighted sum, then..." ,and , , , The specific operating logic is as follows, based on the preset weighting coefficients: S121. Read the normalized parameters of individual fan parameters, real-time wind condition parameters, and environmental parameters; S122, Call the multiplication module to calculate the "normalized fan parameters" respectively. "Normalized wind condition parameters" "and normalized environmental parameters" ”; S123. Call the addition module to sum and obtain the holographic working condition dataset. ; S2, Digital Twin Pre-simulation and Sub-cluster Partitioning, i.e., the holographic working condition dataset received by the digital twin sub-core from the output of S1. Signal: S21. Digital Twin Loading and Clustering Initialization: Calling the 3D Rendering Engine for Loading Rendering engine output The signal generates a 1:1 virtual mapping of the wind field's physical entities; Start the K-means++ clustering algorithm and randomly select An initial cluster center is generated when rendering is complete and the data is received. The signal triggers initial center selection, which follows the rule of covering three wind speed ranges: "low," "medium," and "high." To preset the number of clusters, and ; S22. Sub-cluster iterative partitioning is triggered when the initial cluster centers are determined: S221, Calculate the virtual model of each wind turbine to The Euclidean distance between the initial centers is given by the following formula: ; Among them, wind speed difference represents the difference between the wind speed value corresponding to a certain wind turbine virtual model and the wind speed value of the cluster center; wind turbine type difference is represented as "0" (same type of wind turbine, such as both are new wind turbines within 5 years of service) or "1" (different types of wind turbines, such as new wind turbines and old wind turbines with more than 8 years of service), which is determined according to the preset wind turbine type classification rules; environmental level difference is represented as "0" (same environmental level, such as both are normal temperature and humidity environments) or "1" (different environmental levels, such as normal environment and high salt spray environment), which is determined according to the preset environmental level classification rules; S222. Assign each wind turbine to the nearest cluster center to form a temporary sub-cluster; S223. Recalculate the center value of each temporary sub-cluster. The center value is the average value of all wind turbine parameters in the cluster. S224. Repeat steps S221 to S223 until the Euclidean distance between two adjacent cluster centers is reached. ,in, To preset a clustering error threshold, the error is updated after repeated operations at the center. And output Triggered by a signal; S23, when received The signal triggers the generation of sub-cluster partitioning results, which include cluster number, a list of wind turbines within the cluster, and cluster operating condition labels. .

[0023] S3, Lightweight Federated Learning Aggregation: S31. Global Model Initialization and Lightweighting: When the wind farm control center receives the system startup command, i.e., the startup command sent by the wind farm master controller... At that time, the global federated learning model is initialized. The global federated learning model adopts a recurrent neural network structure, and the input dimension is the same as that of the global federated learning model. The number of parameters is the same, denoted as ; When each wind turbine's AI core receives Download complete signal At that time, initiate lightweight architecture to perform model compression: S311, When loaded into the AI ​​kernel memory, layer pruning is triggered: traverse the model's convolutional layers and remove the absolute values ​​of the weights. The neurons output the pruned model ,in The preset pruning threshold; S312, Received When generating a signal, a bit quantization action is triggered: the high-order floating-point weights are converted to low-order integers, and the quantization rule is as follows: ),in, This is represented as a rounding function; This is represented as a preset quantization coefficient, such as 127, suitable for an 8-bit integer range; the maximum weight value represents the maximum absolute value of the weights of neurons in that layer; the high-order floating-point representation is a 32-bit or 64-bit floating-point number; the low-order integer representation is an 8-bit or 16-bit integer, with a bit width... Output lightweight model And lightweight model Volume is , , The preset model volume threshold is used, and ; S32, AI core received and Local model training is triggered when the synchronization signal is received. S321, from Historical operating data of local wind turbines was extracted as the training set. The historical operating data is recent... Data within the time period, The preset training data duration and sampling interval are: , The preset data sampling interval; S322, with As the initial model, iterative training is performed using a loss function. Each round iterates through the training set once, after each previous round of training has completed and the loss value has decreased. Iterative training is triggered at certain times, where To preset the loss reduction threshold, To preset the number of local training rounds; The loss function is: ; This represents the total number of samples in the training set; Represented as model for the th Predicted wind power or blade angle for each sample; Represented as the first The actual wind power or blade angle value corresponding to each sample; It is expressed as the sum of the squares of the differences between the predicted and actual values ​​for all samples; S323, Round training completed and validation set accuracy Time-triggered output of locally trained model parameters ,in The preset verification accuracy threshold; S33 and AI cores receive the output from S2. With all AI cores within the cluster The signal triggers parameter aggregation within the sub-cluster, and the aggregation action is performed using a federated average aggregation algorithm. The operating logic is as follows: S331. Obtain the preset error statistics period. Read historical fault records collected by S11 and recent The calculation of individual difference weights is triggered when the blade angle error occurs within a certain time period. : Historical failure frequency: ,in, This represents the frequency of wind turbine failures within the statistical period; Represented as The total number of fault events recorded by the wind turbine within the period; Represented as The total actual operating time of the fan within the cycle; Angle error rate: ,in, This is expressed as the relative error rate for wind turbine blade angle control; Represented as The average deviation between the blade angle control command and the actual angle for each cycle; This represents the preset maximum allowable control error for the fan blade angle; The final calculation of individual difference weights: ; in, , For the preset weighting coefficients, and ; The range of values ​​is , , To preset individual weight thresholds, ; S332. When reading the wind speed fluctuation amplitude and turbulence intensity collected by S11, the dynamic wind condition weight is calculated. : Calculation of wind speed fluctuation coefficient: ,in, This is expressed as the relative degree of wind speed fluctuation; the amplitude of wind speed fluctuation is expressed as... The difference between the maximum and minimum wind speeds within a period; This represents the preset maximum wind speed threshold. Calculation of turbulence coefficient: ,in, The relative intensity of turbulence is expressed as ; the intensity of turbulence is expressed as . Standard deviation of wind speed within the period; This is represented as a preset extreme turbulence threshold; The final calculation of dynamic wind condition weights: ,in, , For the preset weighting coefficients, and ; The range of values ​​is , , To preset the wind condition weight threshold, ,and Higher priority ; S333, according to and Calculate the final aggregate weights: ,in, , To preset the final weight coefficients, and ; S334, Aggregate Calculation: ,in, Represented as the first in the cluster Index of typhoon generators; Represented as the first The final aggregate weight of the typhoon generator; Represented as the first Local model parameters output by the typhoon generator AI core; This represents the sum of the final aggregate weights of all wind turbines within the cluster; The output is represented as the sum of the products of all wind turbine model parameters and their corresponding weights within the cluster; the output is the sub-cluster aggregate parameter. ; S34, The wind farm control center receives data from all sub-clusters. Global model aggregation is triggered when a signal is received. S341. Calculate the global weight. The weight of each sub-cluster = the number of wind turbines in the cluster / the total number of wind turbines in the wind farm. S342, Global Aggregation: Wherein, cluster weight is represented as the proportion of the number of wind turbines in a certain sub-cluster to the total number of wind turbines in the wind farm; This can be expressed as the sum of "cluster weight × sub-cluster aggregation parameter" over all sub-clusters; S343, Aggregate Globally Distribute to each AI core.

[0024] In this embodiment, in step S33, the individual difference weighting With dynamic wind condition weights The calculation also includes the following optimization, which is triggered when S11 collects the wind turbine's manufacturing year and wind speed fluctuation amplitude signals for updating: Read the manufacturing year of the fan Triggered Aging compensation: Set the aging assessment period ,like Then increase the mechanical aging compensation coefficient. ,in To preset the aging compensation weight, The formula is adjusted to: ; in, , Same definition as S33; like Then keep the original The formula remains unchanged; Read the wind speed fluctuation range Triggered Fluctuation compensation: Preset high wind speed fluctuation threshold ,like Then increase the wind speed fluctuation compensation coefficient. ,in To preset the fluctuation compensation weight, The formula is adjusted to: ; in, , Same definition as S33; like Then keep the original The formula remains unchanged; According to the compensation , Trigger the final aggregation calculation: Calculate the final weight ; Sub-cluster aggregation parameter calculation: ,in For the first in the cluster Typhoon machines, i.e. all Triggered when calculation is complete; S4. Intelligent resource scheduling and collaborative control execution: S41, Multi-core Resource Allocation: When the PLC of the multi-core processor completes its power-on, the digital twin sub-core initializes an independent hardware thread. The priority of the independent hardware thread is set to "medium," lower than the PLC core but higher than the AI ​​core, and it is dedicated to undertaking the rendering and clustering tasks of S2; when the AI ​​core receives... Time-based blade angle control recommendations ; S42, PLC core control execution: S421, PLC core received PID parameter optimization is triggered when the ambient temperature signal output by S1 is used: Preset number of optimization iterations Inertia weight Learning factors and Maximum overshoot Maximum response time If the temperature sensor detects a temperature Start the particle swarm optimization algorithm and set the number of iterations. Inertia weight Learning factors , ,and ; with overshoot And response time Optimize the PID parameters to the objective function. - , - , - ; in, This is the preset range of optimized scaling factors; This is the preset range of optimized integral coefficients; To preset the range of the optimized differential coefficients, and , , ; To preset the low temperature threshold, and continuously Duration The preset temperature determination time; If temperature If the default PID parameters are used, then... - , - , - ,and , ; When the S422 PLC core receives the wind speed change rate signal output by S1, it triggers extreme operating condition resource reservation: If the rate of change of wind speed ,in To preset the extreme wind speed change rate threshold, and output Signal: S4221. Disable unnecessary training processes for the AI ​​core; S4222, Reserved The proportional PLC computing power, of which To reserve a certain proportion of computing power for extreme operating conditions, and The task scheduler sets the priority of the control task to "highest"; If the rate of change of wind speed Reserved The proportional PLC computing power, of which To reserve a certain proportion of computing power for pre-set normal operating conditions, and ; S423, Control command generation is triggered when PID parameters are determined and computing power reservation is completed: According to Combined with PID parameters, generate blade angle control commands. It is then issued to the wind turbine actuator.

[0025] S5, Closed-loop feedback optimization: S51, the PLC core receives the actual blade angle signal fed back by the fan actuator. Time-triggered error detection and calculation: Acquire actual blade angle signals Recommendations for blade angle control at the corresponding time point The number of sampling points is , The preset number of error sampling points, i.e. Triggered when all sampling points are collected; Calculate the percentage error, point by point. ( to ),in, Represented as the first The actual blade angle value at each sampling point; Represented as the first The AI ​​core control suggested angle value corresponding to each sampling point; Sum and then average: ,in, This represents the total number of sampling points used in error calculation; Indicated as to Sum the absolute values ​​of the angular deviations of each sampling point; This means converting the calculation result into a percentage form; S52, Preset error percentage threshold ,like ,but: PLC generates error data Error data include , , value; Send error data To the AI ​​core; like If no feedback is triggered, the current control command will continue to be executed. S53, AI core receives error data AI kernel model optimization is triggered by the signal: S531, Based on the wind turbine mechanical loss coefficient collected in S11 Trigger the data fusion action, data fusion The formula is: ; in, , To preset the fusion weight coefficients, and ; It is expressed as the mechanical loss coefficient calculated based on the wind turbine's operating time and historical faults; the larger the value, the more severe the loss. S532, Local Model Retraining: To supplement the training set, start The local training round outputs the updated local parameters. Feedback is sent to S33 to participate in the next round of sub-cluster aggregation; in, To pre-set supplementary training rounds, and ; S54. Based on the salt spray concentration signal received by the PLC core from the output of S1. Triggering extreme environment linkage control: When Initiate reinforcement learning Q-table iteration (Q-learning algorithm), and define the state. It is discrete into three levels: "low / medium / high," divided according to a preset concentration range, and the action... "Start high-frequency dehumidification"; Q-value update formula: ,in, Set the learning rate; Preset discount factor; The reward value, i.e. This indicates that the salt spray concentration decreased after dehumidification. This indicates that the concentration has neither decreased nor increased; It is represented as the maximum Q value of all possible actions under the next state s'; After the Q-value iteration is completed and Trigger the high-frequency dehumidification unit: Press Dehumidification is activated periodically, among which, The preset dehumidification cycle and dehumidification duration are: , Preset dehumidification time; When the dehumidification unit starts, it triggers the blade angle adjustment rate, changing the angle adjustment rate from... Down to ,in, The preset normal adjustment rate; To preset the adjustment rate for extreme environments, and To avoid mechanical jamming in salt spray environments; in, The preset salt spray concentration threshold is used, and it continues... Duration The preset salt spray detection time.

[0026] In step S41, a preset wind speed status update cycle is established. Working condition sensing sub-core Wind speed status signal sent at regular intervals The computing power scheduling logic of the PLC with a multi-core processor is triggered at certain times, specifically as follows: when Indicating wind speed And continue When the duration is long enough, the AI ​​core receives the wind speed status signal and changes the local training rounds from... Adjusted to ,in, For the preset round increment, and , To preset the maximum number of training epochs; after the digital twin sub-core receives the signal indicating that the AI ​​core's epoch adjustment is complete, it activates a low-precision rendering mode, reducing the rendering resolution from... Down to , Preset high-precision resolution; To preset low resolution, and ; when instruct And continue During the duration, the AI ​​core remains Local training is performed in rounds; the digital twin sub-core uses standard precision rendering mode with a rendering resolution of [resolution value missing]. ; when Indicating wind speed And continue When the duration is long, the AI ​​core will switch from local training rounds. Down to ,and , The minimum number of training rounds is preset; after receiving the round adjustment signal, the digital twin sub-core disables the rendering function and retains only the clustering calculation. when Indicator of turbulence intensity And continue When the AI ​​core receives an extreme turbulence signal, it pauses local training; the digital twin sub-core pauses rendering and real-time clustering, loads pre-stored historical partitioning results for similar operating conditions, and outputs pre-stored historical sub-cluster partitioning results under extreme turbulence conditions. ;Specifically, It is obtained through clustering calculations performed during digital twin verification, reflecting the sub-cluster partitioning results of the latest operating conditions, and The sub-cluster partitioning results, not calculated in real time but pre-stored in the system, are generated under similar extreme turbulence conditions that have occurred historically. This can be understood as... It is a contingency plan or cache for the system. When the environment is too harsh to perform complex real-time calculations, the system calls this pre-prepared and verified plan.

[0027] in, To preset a low wind speed threshold; The preset wind speed duration is determined. To preset a high wind speed threshold, and ; The preset extreme turbulence threshold; The preset duration for determining the duration of turbulence.

[0028] In step S42, the PLC core receives and The synchronization signal triggers the PLC core's operating condition adaptability verification logic, specifically: Verification dimension parameter calculation: blade angle adjustment amount : ,in, This is represented as the absolute value of the difference between the AI ​​core's suggested angle and the actual current blade angle of the wind turbine; the current blade angle is represented as the value read by the wind turbine blade angle sensor when the PLC verifies it. according to Environmental parameters are used to calculate environmental adaptability. : , in, This is represented as the real-time collected salt spray concentration value; This represents the standard salt spray concentration value under the preset normal operating scenario of the wind farm; This is represented as a real-time atmospheric pressure value; This represents the standard atmospheric pressure value under the preset normal operating scenario of the wind farm; according to Health status of wind turbine parameters calculation equipment : ; in, This represents the preset equipment health statistics period; This is represented by a preset fault impact weighting coefficient; exist , , Trigger threshold determination upon completion of calculation: like or This is triggered when the parameter is determined to be "high risk": setting a verification matching threshold. ; in, A preset low environmental adaptability threshold is set. To preset a low device health threshold; A preset high matching threshold is set. like and This is triggered when the parameter is determined to be "low risk": setting a verification matching threshold. ,in, To preset a low matching threshold, and ; Calculate matching degree : ; in, , , The preset matching degree weight coefficient is used, and ; for The historical optimal adjustment amount under similar operating conditions is stored in the memory; exist The verification result is determined upon completion of the calculation. like :implement ; like : Refusal to execute Send a "re-inference" request to the AI ​​core.

[0029] Example 2 The PLC collaborative control system for wind farm multi-core processors based on digital twins and lightweight federated learning provided by the present invention will be described below. The PLC collaborative control system for wind farm multi-core processors based on digital twins and lightweight federated learning described below can be referred to in correspondence with the PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning described above, and will be further explained in conjunction with Embodiment 1. like Figure 2 The structure shown is... Figure 2 The logic flowchart of the PLC cooperative control system for wind farm multi-core processors based on digital twin and lightweight federated learning provided in the embodiments of this application includes: The multi-core processor PLC module integrates a working condition sensing sub-core, an AI core, a PLC core, and a digital twin sub-core. These sub-cores are physically isolated and configured with independent memory (AI memory, PLC memory, digital twin memory, and a working condition buffer). The independent memory also includes a working condition buffer for storing the holographic working condition dataset output in step S1. The triggering and execution logic of each core is as follows: In this embodiment, the operating condition sensing sub-core includes the following: Wind turbine parameter acquisition unit: The trigger condition is receiving the PARAM_UPDATE signal from the wind turbine controller (i.e., every...). (Timed output), executes the "Read manufacturing year → Retrieve fault log → Calculate wear factor" action, and outputs... ; Wind condition acquisition unit: The trigger condition is receiving data from the wind speed sensor. Signal, i.e., each Frequency output, execute "Sampling wind speed → Calculating turbulence intensity → Output" "action; Environmental acquisition unit: The trigger condition is receiving an image. signal or A timed signal is used to execute the process of "collecting temperature and humidity → detecting salt spray concentration → outputting data". "action; Multimodal fusion unit: Trigger condition is receiving , , Upon receiving the ACK signal, the weighted fusion action in S12 of claim 2 is executed, and the output is... ; In this embodiment, the AI ​​core includes the following: Lightweight compression unit: Trigger condition is receiving of The signal performs a "layer pruning to bit quantization" operation and outputs... ; Local training unit: Triggered by receiving and The synchronization signal executes "loading the training set, calculating the loss function, and iterative training". The "wheel" action outputs... ; Parameter aggregation unit: Trigger condition is receiving and The signal executes the federated average aggregation action of S33 in claim 1, and calculates... , Post-weighted summation, formula Output ; In this embodiment, the PLC core includes the following: Intelligent PID control unit: Trigger condition is receiving and of The signal executes "temperature determination, particle swarm optimization PID, generation". "Action, that is, in" When, iterative formula , ,in, Represented as the first The first particle Dimensional speed; Represented as inertia weight; , Represented as a random number in the interval [0,1]; Represented as the first The optimal position of each individual particle; Represented as the first The first particle The position of the dimension; Represented as the globally optimal position; Operating condition adaptation verification unit: The trigger condition is receiving and Perform the "parameter calculation, threshold determination, and matching degree verification" actions as described in claim 3; Extreme operating condition response unit: The trigger condition is receiving The signal executes "shut down unnecessary AI core processes and reserve..." "Proportional computing power" action; In this embodiment, the digital twin sub-core includes the following: 3D rendering unit: Trigger condition is receiving The LOAD_OK signal executes the "3D engine rendering to generate 1:1 virtual mapping" action, and outputs... ; Intelligent clustering unit: The trigger condition is receiving The signal executes "Kmeans++ initialization, iterative partitioning, cluster allocation, and output". "action; In this embodiment, the wind farm digital twin is deployed in the digital twin sub-core, and the trigger condition is receiving a signal. It outputs virtual mapping and clustering initialization signals, which in turn trigger the intelligent clustering unit; In this embodiment, the lightweight federated learning module is deployed within the AI ​​core, and its trigger condition is receiving a message. and Execute the "Model lightweighting to sub-cluster aggregation" action; In this embodiment, the closed-loop feedback module includes the following: Error calculation unit: Trigger condition is receiving and ,implement" The "point sampling to percentage error calculation" action outputs... ; Data fusion unit: Trigger condition is receiving data. and Perform the fusion action in S53 of claim 1 and output... ; Feedback trigger unit: Triggering condition is ,send To the AI ​​core; In this embodiment, the extreme environment adaptation module includes the following: High-frequency dehumidification unit: Trigger condition is receiving Signal, execute "every Periodic dehumidification "Duration" action; Reinforcement learning control unit: The trigger condition is receiving The signal executes the Q-table iteration action in S54 of claim 1 and outputs a dehumidification trigger signal; The environmental acquisition unit of the working condition sensing sub-core also includes linkage triggering logic: Salt spray concentration sensor: When it detects Sometimes and continuously Duration, outputting the SALT_HIGH signal to the reinforcement learning control unit of the extreme environment adaptation module; When detected Synchronous output (i.e., the same triggering conditions as above) The signal is sent to the extreme condition response unit of the PLC core, triggering the PLC core to adjust the blade adjustment rate from Down to ,in, To preset the normal adjustment rate, Adjust the rate for preset extreme environments; Barometric pressure sensor: When detected When, output The multimodal fusion unit from signal to operating condition sensing subcore; where To preset a low air pressure threshold, and continuously Duration Preset air pressure determination time; The multimodal fusion unit received After the signal (i.e.) Triggered when the signal is valid), the weights of environmental parameters are changed from... Adjusted to ,in ,and The fusion formula is updated as follows: ; in, ; , These are the original preset weighting coefficients; The intelligent clustering unit of the digital twin sub-core also includes cluster optimization logic: Cluster Quantity Detection Unit: Trigger condition is receiving Signals are used to count the number of wind turbines within each cluster. ; Cluster merge is triggered when a cluster exists. Triggered at time Minimum number of wind turbines to preset: 1. Find adjacent operating condition clusters, where adjacent is defined as the difference in wind speed intervals. And the environmental levels are the same. This is a preset threshold for the wind speed difference between clusters, triggered when the cluster tag comparison is completed. 2. The cluster is merged with the adjacent cluster to generate a temporary merged cluster, which is triggered when the merge rule is confirmed. 3. Restart K-means clustering, triggered when temporary cluster merging is generated, and maintain the "neighbor center error" as the iteration termination condition. ", output optimized ; Optimization result delivery: The trigger condition is receiving the result. The signal was replaced with the original one. The parameter aggregation unit sent to the AI ​​core; It also includes a communication module, with the following triggering and execution logic: High-reliability communication protocol unit: Triggering condition 1: The PLC with the multi-core processor establishes a physical connection with the wind farm control center, executes the "protocol handshake to clock synchronization" action, and the clock synchronization error... , The preset synchronization error threshold is used; Triggering condition 2: AI core generation The "parameters packaged to priority flag" action is executed and sent to the wind farm control center; Triggering condition 3: Wind farm control center generated The "model fragmentation to protocol transmission" action is executed and distributed to each AI core; Incremental parameter transmission unit: Triggering condition 1: AI core generation Calculate compared to the previous round Difference ; Triggering condition 2: ,in To preset a small difference threshold, only the difference amount is transmitted. The parameters, where To preset the transmission difference threshold, the amount of data transmitted is the complete parameter. proportion, of which To preset the incremental transmission ratio, and ; Triggering condition 3: This is triggered when the difference is determined to be "significant update", and the transmission is complete. ; The PLC module with a multi-core processor also includes a fault-tolerant unit, and the triggering logic and operating logic are as follows: AI-based fault-tolerant unit: Triggering condition 1: After the AI ​​kernel completes local training, calculate the training loss value. ; Triggering condition 2: ,in To prevent a pre-set high loss threshold from being identified as a training anomaly, the "model rollback to the previous loading iteration" is executed. "action; Triggering condition 3: Restart after rollback is complete. Local training rounds, among which To pre-determine the number of fault-tolerant training rounds, and ; PLC core fault-tolerant unit: Triggering condition 1: PLC core issued Then, start Timeout timer, where The preset instruction timeout period; Triggering condition 2: If no ACK signal is received from the actuator (timeout triggered), the "instruction retransmission" action is executed, and the instruction is retransmitted. Second-rate, The preset number of retransmissions, with an interval between each retransmission. , Preset retransmission interval; Triggering condition 3: If no ACK is received after the second retransmission (i.e., retransmission timeout triggered), output... The signal triggers an emergency shutdown of the wind turbine; This embodiment also includes a log storage module, and the triggering logic and running logic are as follows: Operation log unit: Triggering condition 1: After each core performs a critical action, a log entry is generated. The log entry contains "timestamp, action type, input parameters, and output result". Triggering condition 2: After a log entry is generated, it is written to local non-volatile memory, with a storage period of [duration missing]. , Set the preset log storage period, such as 30 days; Traceability Unit: Trigger condition 1: Every Timed trigger, To preset the traceability upload period, such as The hour, triggered by an hour timer interrupt, reads log entries from local non-volatile memory. Triggering condition 2: Log entry reading is complete, and the entry is uploaded to the traceability system via a trusted node, triggering the entry verification. Verification rule: hash value matching, hash calculation formula. ,in, Represented by the SHA256 hash algorithm, (log entry content) represents the complete information in the log, such as the timestamp, action type, and parameters. Triggering condition 3: After the upload is completed, a traceability identifier ID is generated and associated with the local log, which is triggered when the traceability system returns an upload success signal; It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0030] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning, characterized in that, Includes the following steps: S1. The operating condition sensing sub-core of the PLC with a multi-core processor collects individual wind turbine parameters, real-time wind condition parameters, and environmental parameters of the wind turbines in the wind farm, and performs multi-modal data fusion to generate a holographic operating condition dataset. ; S2, Based on the holographic working condition dataset A virtual mapping of the wind farm's physical entities is constructed using a digital twin sub-core, and a clustering algorithm is used to perform dynamic iterative partitioning of the wind turbines, outputting the sub-cluster partitioning results; S3. Based on the sub-cluster partitioning results, initialize the global federated learning model using the AI ​​kernel and perform lightweight model compression. Within each sub-cluster, use the federated average aggregation algorithm to calculate individual difference weights and dynamic wind condition weights to aggregate parameters. Finally, aggregate the sub-cluster parameters at the global level to generate an updated global model. ; S4. Based on the updated global model and the control suggestions generated by the AI ​​core, computing resources are allocated to the PLC using a multi-core processor, based on the holographic working condition dataset. The system intelligently reserves and schedules computing resources, generates control commands, and sends them to the wind turbine actuators. S5. The error between the actual blade angle and the control suggestion is detected by the PLC core. When the error exceeds the threshold, feedback is triggered to the AI ​​core to optimize the local model. Based on the holographic operating condition dataset, reinforcement learning control is started to adjust the dehumidification and blade angle rate.

2. The PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Upon receiving the preset parameter update cycle sent by the wind turbine controller. Parameter update completion signal output at regular intervals At the same time, individual parameters of each wind turbine in the wind farm are collected, including the year of manufacture, historical fault records, and mechanical loss coefficient; When the wind speed sensor receives each Sampling signal with duration output Real-time wind condition parameters are collected. This includes wind speed fluctuation amplitude and turbulence intensity; among which, Preset wind condition sampling frequency; When changes in environmental parameters are detected Or received When a timed signal is received, environmental parameters are collected, including temperature, humidity, salt spray concentration, and air pressure; among them, The preset parameter change threshold; The preset environmental sampling period; S12. Perform data fusion using weighted fusion rules. First, normalize the individual wind turbine parameters, real-time wind condition parameters, and environmental parameters to the [0,1] interval; then, according to the weight of the individual wind turbine parameters... Real-time wind condition parameter weights Environmental parameter weights "To calculate the weighted sum, then..." ,and , , , The specific operating logic is as follows, based on the preset weighting coefficients: S121. Read the normalized parameters of individual fan parameters, real-time wind condition parameters, and environmental parameters; S122, Call the multiplication module to calculate the "normalized fan parameters" respectively. "Normalized wind condition parameters" "and" normalized environmental parameters ”; S123. Call the addition module to sum and obtain the holographic working condition dataset. .

3. The PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 1 or 2, characterized in that, The specific steps of step S2 are as follows: S21, Call the 3D rendering engine to load. Rendering engine output The signal generates a 1:1 virtual mapping of the wind field's physical entities; the K-means++ clustering algorithm is then activated, and random selection is performed. An initial cluster center is generated when rendering is complete and the data is received. The signal triggers initial center selection, which follows the rule of covering three wind speed ranges: "low," "medium," and "high." To preset the number of clusters, and ; S22. Sub-cluster iterative partitioning is triggered when the initial cluster centers are determined: S221, Calculate the virtual model of each wind turbine to Euclidean distance between the initial centers; S222. Assign each wind turbine to the nearest cluster center to form a temporary sub-cluster; S223. Recalculate the center value of each temporary sub-cluster. The center value is the average value of all wind turbine parameters in the cluster. S224. Repeat steps S221 to S223 until the Euclidean distance between two adjacent cluster centers is reached. Then output Signal, among which, The preset clustering error threshold is used; S23, when received The signal triggers the generation of sub-cluster partitioning results, which include cluster number, a list of wind turbines within the cluster, and cluster operating condition labels. .

4. The PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 3, characterized in that, The specific steps of step S3 are as follows: S31. Obtain the start command sent by the wind farm master controller. When each wind turbine AI core receives Furthermore, when the AI ​​core receives the download completion signal from the buffer, it initiates lightweight architecture to perform model compression and outputs the lightweight model. : S32, from Extract historical operating data of local wind turbines as the training set; As the initial model, iterative training is performed using a loss function. Each round iterates through the training set once, after each previous round of training has completed and the loss value has decreased. Iterative training is triggered at certain times, where To preset the loss reduction threshold, Preset local training rounds; Round training completed and validation set accuracy Time-triggered output of locally trained model parameters ,in The preset verification accuracy threshold; S33, based on the AI ​​core receiving the output from S2. With all AI cores within the cluster The signal triggers parameter aggregation within the sub-cluster, and the aggregation action is performed using a federated average aggregation algorithm to output the sub-cluster aggregated parameters. ; S34, The wind farm control center receives data from all sub-clusters. The signal triggers global model aggregation to obtain the global model. and the global model Distribute to each AI core.

5. The PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 4, characterized in that, The specific execution logic for the aggregation action in step S33 is as follows: S331. Obtain the preset error statistics period. Read historical fault records collected by S11 and recent Blade angle error over a period of time; historical fault frequency determined based on historical fault records. According to recent Blade angle determination angle error rate over time Based on historical failure frequency and angle error rate Calculate individual difference weights ; S332. Calculate dynamic wind condition weights based on the wind speed fluctuation amplitude and turbulence intensity collected in S11. : S333, Weighting based on individual differences With dynamic wind condition weights Calculate the final aggregate weight Based on the final aggregate weight Output sub-cluster aggregation parameters .

6. The PLC cooperative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 4, characterized in that, The specific steps of step S4 are as follows: S41. When the PLC of the multi-core processor completes its power-on, the digital twin sub-core initializes its independent hardware thread; when the AI ​​core receives... Time-based blade angle control recommendations ; S42, According to the PLC core received PID parameter optimization is triggered when the ambient temperature signal output by S1 is used: preset number of optimization iterations. Inertia weight Learning factors and ,and Maximum overshoot Maximum response time If the temperature sensor detects a temperature Initiate the particle swarm optimization algorithm with overshoot. And response time Optimize the PID parameters to the objective function. - , - , - ; in, This is the preset range of optimized scaling factors; This is the preset range of optimized integral coefficients; To preset the range of the optimized differential coefficients, and , , ; To preset the low temperature threshold, and continuously Duration The preset temperature determination time; If temperature If so, the default PID parameters will be used; according to Based on the determined PID parameters, generate blade angle control commands. It is then issued to the wind turbine actuator.

7. The PLC collaborative control method for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 6, characterized in that, The specific steps of step S5 are as follows: S51. Acquire actual blade angle signal Recommendations for blade angle control at the corresponding time point Calculate point by point ( to ),in, Represented as the first The actual blade angle value at each sampling point; Represented as the first The AI ​​core control suggested angle value corresponding to each sampling point; the number of sampling points is... , The preset number of error sampling points is used; the summation is then averaged. The calculation formula is: ,in, This represents the total number of sampling points used in error calculation; Indicated as to Sum the absolute values ​​of the angular deviations of each sampling point; S52, Preset error percentage threshold ,like Then the PLC core generates error data. Send error data To the AI ​​core; if If no feedback is triggered, the current control command will continue to be executed. S53, Based on the wind turbine mechanical loss coefficient collected in S11 Trigger the data fusion action to obtain fused data. ,by To supplement the training set, start The local training round outputs the updated local parameters. Feedback is sent to S33 to participate in the next round of sub-cluster aggregation; among which, To pre-set supplementary training rounds, and ; S54. Based on the salt spray concentration signal received by the PLC core from the output of S1. Triggering extreme environment linkage control: preset salt spray concentration threshold ,when Initiate reinforcement learning Q-table iteration, and after the Q-table iteration is completed and The high-frequency dehumidification unit is triggered at any time by pressing [button]. Dehumidification is activated periodically, among which, The preset dehumidification cycle and dehumidification duration are: , Preset dehumidification time; The reward value, i.e. This indicates that the salt spray concentration decreased after dehumidification. This indicates that the concentration has neither decreased nor increased; When the dehumidification unit starts, it triggers the blade angle adjustment rate, changing the angle adjustment rate from... Down to ,in, The preset normal adjustment rate; To preset the adjustment rate for extreme environments, and To avoid mechanical jamming in salt spray environments.

8. The PLC collaborative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 6, characterized in that, In step S41, a preset wind speed status update cycle is established. Working condition sensing sub-core Wind speed status signal sent at regular intervals The computing power scheduling logic of the PLC with a multi-core processor is triggered at certain times, specifically as follows: when Indicating wind speed And continue When the duration is long enough, the AI ​​core receives the wind speed status signal and changes the local training rounds from... Adjusted to ,in, For the preset round increment, and , To preset the maximum number of training epochs; after the digital twin sub-core receives the signal indicating that the AI ​​core's epoch adjustment is complete, it activates a low-precision rendering mode, reducing the rendering resolution from... Down to , Preset high-precision resolution; To preset low resolution, and ; when instruct And continue During the duration, the AI ​​core remains Local training is performed in rounds; the digital twin sub-core uses standard precision rendering mode with a rendering resolution of [resolution value missing]. ; when Indicating wind speed And continue When the duration is long, the AI ​​core will switch from local training rounds. Down to ,and , The minimum number of training rounds is preset; after receiving the round adjustment signal, the digital twin sub-core disables the rendering function and retains only the clustering calculation. when Indicator of turbulence intensity And continue When the AI ​​core receives an extreme turbulence signal, it pauses local training; the digital twin sub-core pauses rendering and real-time clustering, loads pre-stored historical partitioning results for similar operating conditions, and outputs pre-stored historical sub-cluster partitioning results under extreme turbulence conditions. ; in, To preset a low wind speed threshold; The preset wind speed duration is determined. To preset a high wind speed threshold, and ; The preset extreme turbulence threshold; The preset duration for determining the duration of turbulence.

9. The PLC cooperative control method and system for wind farm multi-core processors based on digital twins and lightweight federated learning according to claim 5, characterized in that, In step S33, individual difference weights With dynamic wind condition weights The calculation also includes the following optimization steps: Read the manufacturing year of the fan Triggered Aging compensation: Preset aging judgment age ,like Then increase the mechanical aging compensation coefficient. , The formula is adjusted to: ; like Then keep the original The formula remains unchanged; Read the wind speed fluctuation range Triggered Fluctuation compensation: Preset high wind speed fluctuation threshold ,like Then increase the wind speed fluctuation compensation coefficient. , The formula is adjusted to: ; like Then keep the original The formula remains unchanged; According to the compensation , Trigger the final aggregation calculation ; According to aggregate calculation Generate sub-cluster aggregation parameters ; in, , , , Preset weighting coefficients; This represents the frequency of wind turbine failures within the statistical period. This is expressed as the relative error rate for wind turbine blade angle control; Preset aging compensation weights; Preset fluctuation compensation weights; The relative degree of fluctuation in wind speed; The relative intensity of turbulence.

10. A PLC cooperative control system for wind farm multi-core processors based on digital twins and lightweight federated learning, characterized in that, include: The PLC module of the multi-core processor integrates a working condition sensing sub-core, an AI core, a PLC core, and a digital twin sub-core. The sensing sub-core, AI core, PLC core, and digital twin sub-core are physically isolated and configured with independent memory. The independent memory also includes a working condition cache area for storing the holographic working condition dataset output in step S1 as described in claim 1. The wind farm digital twin, deployed in the 3D rendering unit of the digital twin sub-core, is used to receive holographic operating condition datasets and output sub-cluster partitioning results to the AI ​​core; The lightweight federated learning module, deployed in the AI ​​kernel, includes the NanoFormer model compression unit and the sub-cluster aggregation unit. After the sub-cluster partitioning results and local model parameters are uploaded, the action from model lightweighting to sub-cluster aggregation is performed. The closed-loop feedback module, connecting the PLC core and the AI ​​core, includes an error calculation unit for receiving the actual blade angle signal as described in S51 of claim 7. Recommendations for blade angle control at the corresponding time point It performs the actions from sampling to MAPE calculation; The data fusion unit is used to receive error data and individual wind turbine parameters, and perform the data fusion action as described in S53 of claim 7. Feedback trigger unit, used to output error data To the AI ​​core; The extreme environment adaptation module includes a high-frequency dehumidification unit for receiving the signal as described in S54 of claim 7. Signal to initiate dehumidification; The reinforcement learning control unit is used to receive... The signal executes the Q-table iteration action in S54 of claim 7.

Images