PLC variable pitch control method and system based on deep learning

Abstract

The embodiment of the application provides a PLC variable-pitch control method and system based on deep learning, which acquires a blade angle calculated by PID and an optimal blade angle generated by DRL in parallel, uses a lightweight confidence network to evaluate the reliability of the DRL decision in real time, dynamically generates a fusion weight, and finally, intelligently weights and fuses the two control instructions based on the weight, and obtains a final blade angle instruction with the aid of a safety envelope. In this way, fine control can be achieved to optimize energy capture and limit load. When the DRL model confidence is low or faces an edge working condition, the system can adaptively and smoothly increase the weight of the PID controller, ensuring the stability and robustness of the control system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a PLC pitch control method and system based on deep learning. Background Technology

[0002] In modern industry, especially in energy systems such as wind power, precise and efficient blade angle control (i.e., pitch control) is crucial for ensuring equipment operating efficiency, stability, and safety. Pitch control systems adjust the tilt angle of wind turbine blades to adapt to different wind speed conditions, thereby optimizing energy capture, limiting load, and ensuring the generator maintains optimal speed. Programmable Logic Controllers (PLCs), as core devices in industrial control, have long been widely used in various automation control tasks due to their superior stability, real-time processing capabilities, and anti-interference characteristics, naturally becoming the implementation platform for most current pitch control solutions.

[0003] However, with the increasing complexity of industrial systems and the growing demand for performance optimization, traditional PLC-based pitch control methods, such as the classic PID controller, while exhibiting good stability and reliability under normal operating conditions, face limitations when dealing with complex, variable, and nonlinear conditions. To overcome these limitations, more advanced control theories, such as deep reinforcement learning (DRL), have been introduced, aiming to achieve more refined and intelligent pitch control through its powerful self-learning and decision-making capabilities. However, deep learning models typically have tens of thousands or even millions of parameters, and their inference latency and memory consumption far exceed the processing capabilities of standard industrial PLCs. Existing technologies impose complex software designed for computationally intensive tasks directly onto simplified hardware designed for high-reliability real-time logic, leading to multiple mismatches at the hardware, software, and system architecture levels. This model lightweighting and PLC deployment issue has become a core bottleneck hindering the widespread application of these advanced technologies in industrial scenarios.

[0004] Therefore, an optimized deep learning-based PLC pitch control method is desired. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a PLC pitch control method and system based on deep learning.

[0006] In a first aspect, embodiments of the present invention provide a PLC pitch control method based on deep learning, comprising: Acquire generator speed setpoint and real-time sensor data; The actual generator speed is extracted from real-time sensor data, and the actual generator speed and generator speed are input into the PID controller deployed in the PLC to obtain the blade angle calculated by the PID controller. Assemble real-time sensor data into a state vector; A lightweight DRL master controller with state vector inputs deployed in the PLC is used to obtain the optimal blade angle; The state vector input is deployed in a lightweight confidence network in the PLC to obtain the confidence fusion weights; Based on the confidence fusion weight, the optimal blade angle and the blade angle calculated by PID are weighted and fused to obtain the mixed blade angle command. A safety envelope is applied to the mixed blade angle command to obtain the final blade angle command.

[0007] Secondly, embodiments of the present invention provide a PLC pitch control system based on deep learning, comprising: The data acquisition module is used to acquire the generator speed setpoint and real-time sensor data; The blade angle calculation module is used to extract the actual generator speed from real-time sensor data and input the actual generator speed and generator speed setpoint into the PID controller deployed on the PLC to obtain the blade angle calculated by the PID controller. The state vector acquisition module is used to assemble real-time sensor data into a state vector; The optimal blade angle control module is used to deploy the state vector input to the lightweight DRL master controller of the PLC to obtain the optimal blade angle. The confidence fusion weight calculation module is used to input the state vector into a lightweight confidence network deployed on the PLC to obtain the confidence fusion weight. The weighted fusion module is used to perform weighted fusion of the optimal blade angle and the blade angle calculated by PID based on the confidence level fusion weight to obtain the mixed blade angle command; The blade angle command generation module is used to perform a safe envelope on the mixed blade angle command to obtain the final blade angle command.

[0008] Compared with existing technologies, the present invention provides a deep learning-based PLC pitch control method and system. This method acquires the blade angle calculated by PID controllers and the optimal blade angle generated by DRL controllers in parallel. It then utilizes a lightweight confidence network to evaluate the reliability of the DRL decisions in real time, dynamically generating fusion weights. Finally, based on these weights, the two control commands are intelligently weighted and fused, supplemented by a safety envelope, to obtain the final blade angle command. This approach enables refined control to optimize energy capture and limit load. When the DRL model confidence is low or when facing edge conditions, the system can adaptively and smoothly increase the weights of the PID controller, ensuring the stability and robustness of the control system. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 A flowchart of a PLC pitch control method based on deep learning according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the data flow of a PLC pitch control method based on deep learning according to an embodiment of the present invention. Figure 3 This is a block diagram of a deep learning-based PLC pitch control system according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0013] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0014] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0015] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0016] In the technical solution of this invention, a PLC pitch control method based on deep learning is proposed. Figure 1 This is a flowchart of a PLC pitch control method based on deep learning according to an embodiment of the present invention. Figure 2 This is a system architecture diagram of a PLC pitch control method based on deep learning according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the deep learning-based PLC pitch control method according to an embodiment of the present invention includes the following steps: S1, acquiring the generator speed setpoint and real-time sensor data; S2, extracting the actual generator speed from the real-time sensor data, and inputting the actual generator speed and the generator speed into the PLC's PID controller to obtain the blade angle calculated by the PID; S3, assembling the real-time sensor data into a state vector; S4, inputting the state vector into the PLC's lightweight DRL master controller to obtain the optimal blade angle; S5, inputting the state vector into the PLC's lightweight confidence network to obtain confidence fusion weights; S6, based on the confidence fusion weights, performing weighted fusion of the optimal blade angle and the blade angle calculated by the PID to obtain a mixed blade angle command; S7, applying a safe envelope to the mixed blade angle command to obtain the final blade angle command.

[0017] Specifically, S1 involves acquiring the generator speed setpoint and real-time sensor data. It should be understood that in modern industrial fields, especially in energy systems such as wind power generation, precise and efficient blade angle control (i.e., pitch control) is crucial for ensuring equipment operating efficiency, stability, and safety. To optimize energy capture, limit load, and ensure the generator maintains optimal speed, the control system must accurately and in real-time grasp its own operating goals and actual state. The generator speed setpoint provides the control system with a clear performance target, directly determining whether the wind turbine pursues maximum wind energy capture efficiency (MPPT mode) or executes power limiting commands. Without this setpoint, the controller will lose its guidance and cannot effectively guide the wind turbine to operate according to the predetermined strategy. Simultaneously, real-time sensor data provides the system's current sensing capabilities, enabling the controller to accurately understand the wind turbine's true physical state under constantly changing natural environments (such as wind speed and direction) and operating conditions (such as actual speed, blade angle, and output power). Only when the controller possesses both a clear target and accurate current status feedback can it effectively calculate deviations, assess the current state, and generate appropriate control commands, thereby achieving high-performance control under complex, variable, and nonlinear operating conditions.

[0018] The generator speed setpoint refers to the target speed value set for the generator based on the current operating strategy (such as the optimal tip speed ratio at a specific wind speed, grid requirements, etc.), serving as a benchmark for measuring control effectiveness. Real-time sensor data, on the other hand, is a multi-dimensional information set encompassing the current system status, including physical quantities such as actual generator speed, wind speed, wind direction, ambient temperature, actual blade angle, and generator output power. These data collectively constitute a comprehensive understanding of the wind turbine system and its operating environment, providing all the necessary decision-making basis for both traditional PID controllers and innovative lightweight DRL main controllers. This data is not only the fundamental input for control but also a key indicator for judging system performance and safety.

[0019] In practice, the generator speed setpoint is typically transmitted to the PLC via an upper-level monitoring system (such as a SCADA system) using industrial communication protocols (e.g., Modbus TCP, Profinet, EtherNet / IP, etc.) and stored in a specific data block within the PLC. The PLC reads this setpoint at the beginning of each control cycle as the target for the current cycle. Real-time sensor data is acquired through various I / O modules connected to the PLC. For example, encoders or speed sensors used to measure the actual generator speed transmit analog or digital signals to the PLC's dedicated high-speed counter module or analog input module; wind speed sensors and wind vanes send wind condition information to the PLC via analog input modules; blade angle encoders (or angle sensors) provide real-time blade angle feedback to the position input module; and various sensors used to monitor generator power, vibration, temperature, and other environmental and operating parameters also collect data that is fed into the PLC. The PLC automatically refreshes the I / O status within each scan cycle, converting these physical quantities into digital signals and storing them in its internal Process Image Input (PII) area, ensuring the control program has access to the most up-to-date data.

[0020] Specifically, in step S2, the actual generator speed is extracted from real-time sensor data, and the actual generator speed and the generator speed are input to the PID controller deployed in the PLC to obtain the blade angle calculated by the PID controller. In the technical solution of this invention, by allowing the PID controller to run continuously and output a blade angle command, the system not only provides a quantifiable and robust reference decision for subsequent intelligent integration, but also ensures that under any extreme or unknown conditions, when the confidence of the DRL controller is insufficient, the system can seamlessly return to a stable control domain that has been verified by long-term industrial practice, thereby providing a fundamental guarantee for the robustness and security of the entire hybrid intelligent system.

[0021] In practice, the first step is to calculate the speed error between the actual generator speed and the generator speed setpoint. This process begins by accurately extracting the actual generator speed from the sensor data set obtained in the previous step. This is achieved by directly reading specific input variables from a PLC programming language (such as ladder diagrams or structured text). These variable values ​​are updated in real-time in the PLC's internal memory (such as the process image input area) after the physical signals are converted into digital quantities by the PLC I / O module. Next, the speed error between the actual generator speed and the generator speed setpoint is calculated. This error value quantifies the degree and direction of the current generator speed deviating from the target value.

[0022] Then, the speed error is input to the PID controller deployed in the PLC to obtain the blade angle calculated by the PID controller. In the PLC programming environment, the standard PID control algorithm is usually encapsulated in the form of a pre-built function block (FB) or a custom program module. This function block receives the speed error calculated in the previous step as input and performs internal calculations based on preset proportional (P), integral (I), and derivative (D) gain parameters. For example, the proportional term generates a corresponding output based on the current error magnitude; the integral term eliminates the steady-state error of the system (i.e., the deviation that exists for a long time); and the derivative term reacts to the rate of change of the error to reduce overshoot and speed up the system response. After these calculations, the PID function block outputs a blade angle value representing the PID controller's recommendation at the current moment. This blade angle is a pitch command generated by the PID controller based on the current speed error, which aims to adjust the rotor torque by changing the angle of attack between the blades and the wind, thereby enabling the generator speed to quickly and accurately approach the set point.

[0023] Specifically, S3 assembles real-time sensor data into a state vector. That is, it transforms the raw physical quantities collected from various sensors, each with different formats and units, into a standardized, multi-dimensional form that can be used by a deep learning model. Here, it should be understood that unlike traditional PID controllers which only require a single speed error as input, the powerful self-learning and decision-making capabilities of deep reinforcement learning rely on a comprehensive and complete description of the current system state. To effectively handle complex, variable, and nonlinear operating conditions, the model must simultaneously consider multiple interrelated variables such as wind speed, generator speed, current blade angle, and output power. Therefore, integrating these scattered data into a structured state vector is the only way to provide the complete contextual information needed for decision-making. The state vector is an ordered, typically normalized, numerical array, where each element represents a key state feature of the system at a specific moment, representing a structured and standardized abstract representation of the system state.

[0024] In practice, the PLC program first selectively extracts relevant variables from the real-time sensor data set acquired in the previous step, based on predefined model input requirements. This is not a simple data listing, but a purposeful filtering process to ensure that only the most valuable information for pitch control decisions is considered. Next, these selected, independent variable values ​​are arranged in a fixed order and stored in a contiguous memory area, forming an array or structure with fixed dimensions and order. For example, the program might create a floating-point array, storing the actual generator speed in the first element, the current wind speed in the second, and so on. Furthermore, because the physical units and numerical ranges of the original sensor data vary greatly, directly inputting them into the neural network would lead to training difficulties and decreased model performance. Therefore, each element in the vector must be normalized, scaling it to a uniform, smaller interval (such as 0 to 1 or -1 to 1). A commonly used normalization method is min-max normalization. In this method, the maximum and minimum values ​​of each variable are constants pre-defined based on the design and operating parameters of the wind turbine generator and are fixed in the PLC program. After normalization, the state vector is obtained.

[0025] Specifically, in S4, the state vector input is deployed to the lightweight DRL master controller of the PLC to obtain the optimal blade angle. It should be understood that while traditional PID controllers are stable, they often fall short in handling complex, variable, and nonlinear operating conditions. Therefore, in the technical solution of this invention, more advanced control theories, such as deep reinforcement learning (DRL), are introduced to achieve more refined and intelligent pitch control through their powerful self-learning and decision-making capabilities. The lightweight DRL master controller, through nonlinear mapping and inference of the comprehensive state vectors, can learn and execute the optimal control strategy under specific operating conditions. Its goal is not merely to stabilize the rotational speed like a PID controller, but also to comprehensively consider multiple objectives such as optimizing energy capture and limiting load while stabilizing the rotational speed, thereby finding a globally optimal blade angle. Ultimately, a highly intelligent control decision oriented towards performance optimization is generated, thereby improving the overall system's operating efficiency and economic benefits.

[0026] The lightweight DRL master controller refers to a specially designed and optimized deep reinforcement learning model. "Lightweight" means that through techniques such as model compression, knowledge distillation, and quantization, it has fewer parameters and lower computational complexity. This addresses the core bottleneck of deep learning models, which typically have tens of thousands or even millions of parameters, resulting in inference latency and memory consumption far exceeding the processing capabilities of standard industrial PLCs. It enables fast, real-time forward inference on resource-constrained PLCs. DRL, on the other hand, is a machine learning method that allows an agent to learn how to make decisions through interaction with its environment to maximize cumulative rewards.

[0027] In practice, the PLC program first loads the normalized state vector generated in the previous step into a designated memory region, which is the input layer of the lightweight DRL master controller neural network model. The size of this input layer perfectly matches the dimension of the state vector. Next, the PLC calls a function library or custom code specifically for performing neural network calculations. This process follows the hierarchical structure of the neural network, starting from the input layer and performing calculations layer by layer until the output layer. At each layer (e.g., a fully connected layer), a combination of linear transformation (matrix multiplication) and nonlinear activation (such as the ReLU function) is performed. This calculation process is passed layer by layer until the last layer of the model, the output layer. Furthermore, the output layer of the DRL master controller typically has only one neuron, whose output value is a scalar located in the normalized interval (e.g., 0 to 1 or -1 to 1). This scalar value represents the blade angle that the model considers optimal, but is still normalized. Therefore, it is further denormalized to convert it back to actual physical angle units to obtain the optimal blade angle. Here, the optimal blade angle is the blade angle setting value that achieves the best overall performance under the current operating conditions.

[0028] Specifically, in S5, the state vector input is deployed in a lightweight confidence network of the PLC to obtain confidence fusion weights. It should be understood that while deep learning models are powerful, their decisions are not entirely reliable, especially when facing edge cases that are not present or insufficiently covered in the training data, where their performance may degrade. Simply or in a fixed way, fusing the output of the DRL controller with the output of the PID controller may introduce instability in high-risk scenarios. Therefore, in the technical solution of this invention, a parallel lightweight confidence network is used to evaluate the reliability of the DRL master controller's decisions under the current operating condition in real time. This confidence network can determine, based on the current state vector, whether the operating condition falls within the range that the DRL model is familiar with and can make high-quality decisions. This gives the entire control system adaptive capabilities, allowing it to fully utilize its optimization performance when the DRL decision is reliable, and automatically favoring more conservative and robust PID control when the DRL decision is questionable, thereby achieving a dynamic balance between pursuing optimal performance and ensuring system safety.

[0029] The lightweight confidence network is a small neural network that runs in parallel with the DRL master controller.

[0030] In practice, firstly, the PLC program uses the state vector as input to the lightweight confidence network to ensure that decisions and evaluations are based on the same operating condition information. Then, similar to the DRL master controller, the PLC calls the neural network inference function to perform a complete forward propagation calculation on the confidence network. This network also consists of multiple layers, operating through a series of linear transformations and nonlinear activation functions. The output layer of the confidence network is typically designed as a single neuron, and its final activation function is usually the Sigmoid function. The Sigmoid function maps any real-valued input to the (0, 1) interval, ultimately yielding the confidence fusion weights.

[0031] Specifically, in S6, the optimal blade angle and the blade angle calculated by the PID controller are weighted and fused based on confidence level to obtain a mixed blade angle command. It should be understood that the linear weighted fusion mechanism fails to fully utilize the key information inherent in the difference between the output values ​​of the deep learning (DRL) controller and the traditional PID controller. Specifically, it ignores the important special relationship of decision divergence between the controllers. When the commands output by the two controllers differ significantly, this is itself a strong warning signal, suggesting that the system may be in a highly nonlinear, transient, or insufficiently learned edge condition, at which point the decision risk of the DRL controller increases significantly. The weights of the existing fusion mechanism are determined only by the internal confidence level of the DRL model, without considering the external consistency between its decisions and the robust PID benchmark. This blind spot may lead to the system still assigning a large weight to risky decisions even in scenarios with high confidence and large divergence, potentially jeopardizing system safety. To address the aforementioned technical deficiencies, this invention proposes a preferred example, which uses a dual-modulation dynamic fusion method incorporating a decision divergence attenuation factor to weightedly fuse the optimal blade angle and the blade angle calculated by PID control to obtain a mixed blade angle command.

[0032] In practical implementation, firstly, based on the optimal blade angle and the blade angle calculated by the PID controller, the decision divergence degree is determined. That is, a clear mathematical metric is used to objectively measure the degree to which the DRL controller's decision deviates from the robust baseline of the PID controller, providing a reliable and timely input for subsequent risk assessment and weight modulation. During execution, the system acquires the optimal blade angle calculated by the DRL controller within one control cycle. Compared with the reference blade angle calculated by the PID controller The absolute difference between the two is calculated and used as the degree of decision divergence. It successfully transforms the abstract differences between two control philosophies into a concrete engineering scalar that can be calculated in real time within the PLC. The value of this scalar directly reflects the aggressiveness or potential risk of the DRL decision under the current operating conditions. This provides a quantitative indicator that can instantly characterize the risk of system decisions, laying the data foundation for truly intelligent and safe dynamic allocation of control. Specifically, this process is expressed by the formula:

[0033] in, For the optimal blade angle, The blade angle is calculated by PID control.

[0034] Next, based on the decision divergence degree and divergence sensitivity coefficient, a divergence attenuation factor is determined. That is, the original, linear divergence degree signal is converted into a non-linear modulation factor with a penalty effect, ensuring that small, normal-range decision divergences do not affect system performance, while significant, dangerous decision divergences can decisively suppress the control authority of the DRL. During execution, the decision divergence degree calculated in the previous step is... Substituting a preset exponential decay function, a bifurcation decay factor ranging between 0 and 1 is generated. This step constructs a dynamic safety valve. When the two controllers make the same decision, the valve is fully open (attenuation factor of 1), without affecting the optimal performance of the DRL. However, when the decision divergence increases, the valve closes rapidly in an exponential manner (attenuation factor approaching 0), forcibly weakening the DRL's control dominance. This establishes an adaptive penalty mechanism highly sensitive to decision risk, ensuring that the system automatically and smoothly increases its conservatism when facing potential uncertainty. Specifically, this process is expressed by the following formula:

[0035] in, The divergence decay factor represents the period t; This represents a bifurcation sensitivity coefficient used to adjust the degree of attenuation, which is a preset normal value; This represents the decision disagreement degree calculated in the previous step.

[0036] Furthermore, based on the confidence fusion weights and the divergence decay factor, the optimal blade angle and the blade angle calculated by the PID are fused to obtain the hybrid blade angle command. That is, the internal self-confidence of the DRL model is organically combined with the divergence decay factor based on external decision consistency to form a more comprehensive and robust final control weight. During execution, the original confidence weights output by the confidence network are first... The divergence decay factor generated in the previous step Multiplying these results in a final effective weight that has undergone double modulation. Subsequently, this final effective weight is used to perform a weighted summation of the outputs of the DRL controller and the PID controller to obtain the final fused blade angle command. This ensures that the DRL controller can only gain primary control of the system under the dual conditions of high internal confidence and reasonable external performance. This process achieves an optimal balance: it fully unleashes the potential of the DRL controller to pursue optimal performance under normal operating conditions, while seamlessly and forcefully handing over control to the more robust PID controller under any unknown or extreme conditions through a divergence decay mechanism, thus generating a final control command that balances optimality and safety. Specifically, this process is expressed by the following formula:

[0037]

[0038] in, The final effective weight representing period t; γ_t represents the original confidence weights; γ_t represents the divergence decay factor. The final fusion blade angle for period t; and The definition is the same as before.

[0039] In summary, this preferred example, by introducing decision divergence as a second-dimensional modulation factor, fundamentally overcomes the single-dimensional fusion deficiency of the original mechanism, which relies solely on the self-evaluation of the DRL model. Ultimately, an advanced hybrid control architecture capable of achieving a dynamic adaptive balance between performance and safety is constructed. This architecture not only fully leverages the optimal performance of deep learning models in improving power generation efficiency and suppressing loads under normal operating conditions, but also, when the system faces edge conditions or potential risks not covered by the model, automatically, quickly, and smoothly regresses to the stability domain of the PID controller, which has been validated through long-term industrial practice, based on the decision divergence between controllers. This significantly enhances the robustness and safety of the entire control system across the entire operating range, achieving a deep synergy and essential fusion of the optimality of artificial intelligence and the reliability of classical control theory.

[0040] Specifically, in S7, a safety envelope is applied to the mixed blade angle command to obtain the final blade angle command. It should be understood that although the front-end hybrid control strategy has greatly enhanced the robustness of the system through the introduction of a PID controller and arbitration by a confidence network, there is still a slight possibility, theoretically, that the mixed blade angle command might exceed the safe operating range allowed by the mechanical structure and aerodynamics of the wind turbine due to model defects, transient sensor failures, or unforeseen combinations of extreme operating conditions. For example, the command might require the blade to change at an excessively fast rate, exceeding the execution capability of the hydraulic or electric pitch control system, thus causing equipment damage; or the command might make the blade angle too small or too large, triggering stall or structural overload. Therefore, in the technical solution of this invention, by applying a safety envelope to the mixed blade angle command, a mandatory final verification and restriction is performed on the command output by the front-end algorithm, ensuring that the final blade angle command issued to the actuator is always within a preset, absolutely safe operating range. This avoids any control behavior that might endanger equipment safety or lead to catastrophic accidents, providing a final and indispensable safety guarantee for the industrial application of the entire intelligent control system.

[0041] In practice, the first step is to compare the received mixed blade angle command with preset upper and lower limits for the blade angle. These two limits (e.g., minimum blade angle and maximum blade angle) are fixed values ​​determined based on the design specifications and aerodynamic characteristics of the wind turbine generator, representing the absolute range of movement that the blades can physically achieve. The logic can be implemented using a simple conditional statement: if the command value is greater than the preset maximum blade angle, then the command value is forcibly set to this maximum value; conversely, if the command value is less than the preset minimum blade angle, then it is forcibly set to this minimum value. Next, to prevent impact and wear on the pitch actuator, the program also limits the rate of change of the blade angle. During this process, the difference between the current command and the final command of the previous cycle is calculated and compared with the preset maximum allowable pitch rate to obtain the required rate of change. If the absolute value of the calculated rate of change exceeds the preset maximum allowable pitch rate, then the current command will be corrected. The revised calculation logic is to add or subtract a maximum permissible change determined by the maximum permissible pitch rate and control cycle time, based on the final instruction of the previous cycle, to obtain a new instruction value. The instruction obtained after these two constraints is the final blade angle instruction. The final blade angle instruction is the final control signal that the PLC ultimately sends to the pitch driver through its output module and is executed by the physical actuator, ensuring absolute safety.

[0042] In summary, the deep learning-based PLC pitch control method according to embodiments of the present invention is explained. It acquires the blade angle calculated by the PID controller and the optimal blade angle generated by the DRL controller in parallel, and utilizes a lightweight confidence network to evaluate the reliability of the DRL decision in real time, dynamically generating fusion weights. Finally, based on these weights, the two control commands are intelligently weighted and fused, supplemented by a safety envelope, to obtain the final blade angle command. In this way, fine-grained control can be achieved to optimize energy capture and limit load. When the DRL model confidence is low or when facing edge conditions, the system can adaptively and smoothly increase the weights of the PID controller, ensuring the stability and robustness of the control system.

[0043] Furthermore, a PLC pitch control system based on deep learning is also provided.

[0044] Figure 3 This is a block diagram of a PLC pitch control system based on deep learning according to an embodiment of the present invention. Figure 3 As shown, the deep learning-based PLC pitch control system 300 according to an embodiment of the present invention includes: a data acquisition module 310 for acquiring generator speed setpoint and real-time sensor data; a blade angle calculation module 320 for extracting the actual generator speed from the real-time sensor data and inputting the actual generator speed and generator speed setpoint into a PID controller deployed in the PLC to obtain the blade angle calculated by the PID; a state vector acquisition module 330 for assembling real-time sensor data into a state vector; an optimal blade angle control module 340 for inputting the state vector into a lightweight DRL main controller deployed in the PLC to obtain the optimal blade angle; a confidence fusion weight calculation module 350 for inputting the state vector into a lightweight confidence network deployed in the PLC to obtain confidence fusion weights; a weighted fusion module 360 ​​for weighted fusion of the optimal blade angle and the blade angle calculated by the PID based on the confidence fusion weights to obtain a mixed blade angle command; and a blade angle command generation module 370 for performing a safe envelope on the mixed blade angle command to obtain the final blade angle command.

[0045] Furthermore, the blade angle calculation module 320 is specifically used to: calculate the speed error between the actual generator speed and the generator speed setpoint; and input the speed error into the PID controller deployed in the PLC to obtain the blade angle calculated by the PID.

[0046] Furthermore, the weighted fusion module 360 ​​is specifically used to: determine the decision divergence degree based on the optimal blade angle and the blade angle calculated by PID; determine the divergence attenuation factor based on the decision divergence degree and the divergence sensitivity coefficient; and fuse the optimal blade angle and the blade angle calculated by PID based on the confidence fusion weight and the divergence attenuation factor to obtain the mixed blade angle command.

[0047] As described above, the deep learning-based PLC pitch control system 300 according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with deep learning-based PLC pitch control algorithms. In one possible implementation, the deep learning-based PLC pitch control system 300 according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the deep learning-based PLC pitch control system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the deep learning-based PLC pitch control system 300 can also be one of many hardware modules of the wireless terminal.

[0048] Alternatively, in another example, the deep learning-based PLC pitch control system 300 and the wireless terminal can also be separate devices, and the deep learning-based PLC pitch control system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0049] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A PLC pitch control method based on deep learning, characterized in that, include: Acquire generator speed setpoint and real-time sensor data; The actual generator speed is extracted from real-time sensor data, and the actual generator speed and generator speed setpoint are input into the PID controller deployed on the PLC to obtain the blade angle calculated by the PID controller. Assemble real-time sensor data into a state vector; A lightweight DRL master controller with state vector inputs deployed in the PLC is used to obtain the optimal blade angle; The state vector input is deployed in a lightweight confidence network in the PLC to obtain the confidence fusion weights; Based on the confidence fusion weight, the optimal blade angle and the blade angle calculated by PID are weighted and fused to obtain the mixed blade angle command. A safety envelope is applied to the mixed blade angle command to obtain the final blade angle command.

2. The PLC pitch control method based on deep learning according to claim 1, characterized in that, The actual generator speed is extracted from real-time sensor data, and the actual generator speed and generator speed setpoint are input into a PID controller deployed in the PLC to obtain the blade angle calculated by the PID controller, including: Calculate the speed error between the actual generator speed and the generator speed setpoint; The speed error is input into the PID controller deployed in the PLC to obtain the blade angle calculated by the PID.

3. The PLC pitch control method based on deep learning according to claim 1, characterized in that, Based on confidence-based fusion weights, the optimal blade angle and the blade angle calculated by PID are weighted and fused to obtain the hybrid blade angle command, including: The degree of decision divergence is determined based on the optimal blade angle and the blade angle calculated by PID. Determine the divergence attenuation factor based on the degree of decision divergence and the divergence sensitivity coefficient; Based on the confidence fusion weight and the divergence attenuation factor, the optimal blade angle and the blade angle calculated by PID are fused to obtain the mixed blade angle command.

4. The PLC pitch control method based on deep learning according to claim 3, characterized in that, Based on the optimal blade angle and the blade angle calculated by PID, the decision divergence degree is determined, including: determining the decision divergence degree using the following formula, where the formula is: in, For the optimal blade angle, The blade angle is calculated by PID control.

5. The PLC pitch control method based on deep learning according to claim 3, characterized in that, Based on the degree of decision disagreement and the disagreement sensitivity coefficient, a disagreement attenuation factor is determined, including: determining the disagreement attenuation factor using the following formula, wherein the formula is: in, To determine the degree of disagreement in decision-making, This represents the disagreement sensitivity coefficient.

6. The PLC pitch control method based on deep learning according to claim 3, characterized in that, Based on the confidence fusion weight and the divergence attenuation factor, the optimal blade angle and the blade angle calculated by the PID are fused to obtain the hybrid blade angle command. This includes fusing the optimal blade angle and the blade angle calculated by the PID using the following formula: in, For confidence level fusion weights, As the divergence attenuation factor, For the optimal blade angle and The blade angle is calculated by PID control.

7. A PLC pitch control system based on deep learning, characterized in that, include: The data acquisition module is used to acquire the generator speed setpoint and real-time sensor data; The blade angle calculation module is used to extract the actual generator speed from real-time sensor data and input the actual generator speed and generator speed setpoint into the PID controller deployed on the PLC to obtain the blade angle calculated by the PID controller. The state vector acquisition module is used to assemble real-time sensor data into a state vector; The optimal blade angle control module is used to deploy the state vector input to the lightweight DRL master controller of the PLC to obtain the optimal blade angle. The confidence fusion weight calculation module is used to input the state vector into a lightweight confidence network deployed on the PLC to obtain the confidence fusion weight. The weighted fusion module is used to perform weighted fusion of the optimal blade angle and the blade angle calculated by PID based on the confidence level fusion weight to obtain the mixed blade angle command; The blade angle command generation module is used to perform a safe envelope on the mixed blade angle command to obtain the final blade angle command.

8. The PLC pitch control system based on deep learning according to claim 7, characterized in that, The blade angle calculation module is further used for: Calculate the speed error between the actual generator speed and the generator speed setpoint; The speed error is input into the PID controller deployed in the PLC to obtain the blade angle calculated by the PID.

9. The deep learning-based PLC pitch control system according to claim 7, characterized in that, The weighted fusion module is further used for: The degree of decision divergence is determined based on the optimal blade angle and the blade angle calculated by PID. Determine the divergence attenuation factor based on the degree of decision divergence and the divergence sensitivity coefficient; Based on the confidence fusion weight and the divergence attenuation factor, the optimal blade angle and the blade angle calculated by PID are fused to obtain the mixed blade angle command.

Images