Wind turbine adaptive life extension control method based on reinforcement learning

Abstract

The application relates to the field of life extension control, and discloses a wind turbine adaptive life extension control method based on reinforcement learning, which is used for providing reliable technical means for the life extension of a wind turbine. The method comprises the following steps: collecting the operation state of the wind turbine and key component health parameters, generating multi-dimensional state information; establishing a damage evolution model according to the multi-dimensional state information, predicting the residual life and generating a health degree index; designing a multi-objective optimization strategy, solving and generating a cooperative control instruction; and after the execution of the instruction, dynamically updating the model and the strategy based on the wind turbine state feedback, so as to form a closed-loop control system. The application effectively prolongs the life of the wind turbine.

Description

Technical Field

[0001] This invention relates to the field of life extension control, and more particularly to an adaptive life extension control method for wind turbines based on reinforcement learning. Background Technology

[0002] With the continued growth of global demand for clean energy, wind power, as an important component of the renewable energy sector, is experiencing continuous expansion in installed capacity and power generation scale. Wind turbines, as the core equipment of wind power systems, operate in complex and ever-changing natural environments, enduring the combined effects of wind loads, gravity loads, temperature variations, and other factors. Key components such as blades, gearboxes, and generators are highly susceptible to fatigue damage and performance degradation, which in turn affects the safe and stable operation of the unit, power generation efficiency, and shortens its service life. Therefore, how to effectively extend the service life of wind turbines and improve their operational reliability and economy has become a critical issue that urgently needs to be addressed in the current wind power industry.

[0003] Traditional wind turbine control methods primarily focus on optimizing power generation efficiency by adjusting parameters such as pitch angle and generator torque to achieve maximum power point tracking (MPPT) and improve wind energy conversion efficiency. However, these methods often neglect fatigue load control of critical components. In pursuing high power generation efficiency, this may lead to excessive fatigue damage accumulation in critical components, accelerating performance degradation and shortening the turbine's lifespan.

[0004] Traditional control methods mainly focus on optimizing power generation efficiency, neglecting fatigue load control of key components. This results in the unit pursuing high power generation efficiency while key components suffer excessive fatigue damage, accelerating performance degradation and shortening service life. Existing health monitoring and assessment methods mostly rely on single sensor data or simple empirical models, which are difficult to comprehensively and accurately reflect the actual health status and remaining life of key components, and cannot provide a reliable basis for the formulation of control strategies. Existing control strategies are mostly open-loop control, lacking real-time feedback and dynamic adjustment mechanisms for control effects. This makes it difficult to adapt to dynamic changes in unit operating status and the challenges of harsh operating conditions, resulting in poor control performance.

[0005] Therefore, we propose a reinforcement learning-based adaptive life extension control method for wind turbines to address the above problems. Summary of the Invention

[0006] This invention provides an adaptive lifespan extension control method for wind turbines based on reinforcement learning, which provides a reliable technical means to extend the lifespan of wind turbines.

[0007] The first aspect of this invention provides a reinforcement learning-based adaptive lifespan extension control method for wind turbines. The method includes: collecting wind turbine operating state parameters and key component health status parameters to generate multi-dimensional state information, including real-time operating data and health assessment results; establishing a damage evolution model based on the multi-dimensional state information, and generating remaining lifespan prediction results and health indicators for key components based on the damage evolution model; designing a multi-objective optimization strategy based on the remaining lifespan prediction results and health indicators, including power generation efficiency optimization and fatigue load control objectives; solving the multi-objective optimization strategy to generate coordinated control commands for coordinating pitch and torque actions; executing the coordinated control commands, and dynamically updating the damage evolution model and optimization strategy based on the turbine state feedback after command execution, forming a lifespan extension closed-loop control system.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the method includes: acquiring wind speed, rotational speed, power, pitch angle, and generator torque parameters during the operation of the wind turbine in real time, and generating a real-time operation dataset; synchronously acquiring vibration signals, temperature signals, and structural strain signals of key components, and generating an original health monitoring dataset; performing time-frequency domain analysis on the vibration signals in the original health monitoring dataset, extracting characteristic frequency components and vibration energy distribution, and generating vibration characteristic indicators; performing trend analysis and thermal load assessment on the temperature signals in the original health monitoring dataset to generate temperature health indicators; performing rainflow counting and load spectrum analysis on the structural strain signals in the original health monitoring dataset, calculating the cumulative fatigue damage, and generating health indicators; and fusing the real-time operation dataset, vibration characteristic indicators, temperature health indicators, and health indicators to generate multidimensional state information.

[0009] Optionally, in a second implementation of the first aspect of the present invention, the method includes: constructing a physical failure mechanism model based on the material properties, geometric configuration, and design load spectrum of the key components, and generating a basic failure model; using real-time operating data and health indicators from the multi-dimensional state information, calculating the equivalent load and stress cycle borne by the key components under the current operating conditions, and generating a real-time load spectrum; inputting the real-time load spectrum into the basic failure model, calculating the damage increment caused by the current operating state, and superimposing it with historical accumulated damage to generate an updated damage accumulation state; Based on the updated damage accumulation state and the failure threshold defined in the basic failure model, the running time or number of cycles required to reach the failure threshold under a preset future operating scenario is estimated, generating a preliminary remaining life prediction value. The preliminary remaining life prediction value is then corrected by combining the vibration characteristic index and the temperature health index, and a health index is generated by integrating the damage accumulation state and the corrected remaining life.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the method includes: dynamically determining the priority relationship between the power generation efficiency optimization objective and the fatigue load control objective based on the numerical range of the health index and the urgency of the remaining life prediction result, and generating dynamic weighting coefficients; establishing a power generation efficiency optimization objective function with the goal of maximizing power generation based on the wind speed and power data in the real-time operating data; establishing a fatigue load control objective function with the goal of minimizing the cumulative fatigue damage rate of key components based on the health index and the real-time load spectrum; integrating the power generation efficiency optimization objective function and the fatigue load control objective function, and using the dynamic weighting coefficients to weight and fuse the two to generate a comprehensive objective function; setting operational safety constraints for the pitch angle, generator torque, and stress of key components based on the design safety boundary of key components and the current operating conditions; and combining and encapsulating the comprehensive objective function and the operational safety constraints to generate a multi-objective optimization strategy.

[0011] Optionally, in the fourth implementation of the first aspect of the present invention, after designing the multi-objective optimization strategy, the following steps are performed: based on the real-time load spectrum and the material SN curve or crack propagation rate curve of the key components, the theoretical fatigue damage threshold of each component under the current operating conditions is calculated, and component-level load safety boundaries are generated. Based on the health index and remaining life prediction results, the load safety boundary of components in the early warning state is dynamically lowered to generate a dynamic load safety boundary. This dynamic load safety boundary is then used as a new hard constraint to supplement and replace some of the original operational safety constraints in the multi-objective optimization strategy, generating an updated multi-objective optimization strategy with integrated health state constraints. When using an adaptive search algorithm to solve the problem, priority is given to ensuring that the generated candidate control action sequences satisfy the dynamic load safety boundary, thereby guiding the search process towards a control region more favorable to vulnerable components, generating a health-oriented search space. The rationality of the dynamic load safety boundary is verified using the feedback from the execution of the cooperative control commands obtained in the health-oriented search space, and fine-tuned based on the verification results to generate a dynamic load safety boundary library.

[0012] Optionally, in the fifth implementation of the first aspect of the present invention, the step of using an adaptive search algorithm to solve the multi-objective optimization strategy and generate a coordinated control command for coordinating pitch and torque actions includes: based on the current multi-dimensional state information and the multi-objective optimization strategy, initializing a set of candidate control action sequences containing a combination of pitch angle and generator torque to generate an initial action population; performing a feasibility check on each action sequence in the initial action population according to the operational safety constraints, eliminating sequences that do not meet the constraints, and generating a feasible action population; using the comprehensive objective function to perform performance calculations on each action sequence in the feasible action population, generating a performance evaluation result; adjusting the search direction and step size according to the performance evaluation result, and generating a new generation of action population through crossover and mutation operations; repeating the process until the performance improvement of the action population is less than a preset threshold or the maximum number of iterations is reached, generating an optimized action sequence; selecting the action combination with the optimal comprehensive objective function value from the optimized action sequence, extracting its corresponding pitch angle setpoint and generator torque setpoint, and generating a coordinated control command for the current control cycle.

[0013] Optionally, in a sixth implementation of the first aspect of the present invention, the step of executing the coordinated control command and dynamically updating the damage evolution model and optimization strategy based on the unit status feedback after the command execution to form a life extension closed-loop control system includes: sending the coordinated control command to the pitch system and generator torque control system of the wind turbine and executing it to generate corresponding actual control actions; after executing the coordinated control command, collecting the wind turbine operating status parameters and key component health status parameters for a new monitoring cycle to generate unit status feedback data; based on the unit status feedback data, analyzing the actual impact of the coordinated control command on the fatigue damage accumulation rate of key components and power generation efficiency, and generating a control effect evaluation report; and according to the control effect... Based on the evaluation report, it is determined whether the actual damage evolution trend of the key components is consistent with the predicted trend of the damage evolution model. The relevant parameters in the damage evolution model are calibrated according to the deviation to generate an updated damage evolution model. The updated damage evolution model is then used to reassess the health status of the key components, generating updated health indicators and remaining life prediction results. Based on the updated health indicators and remaining life prediction results, the weight coefficients of the objective function or the operational safety constraints in the multi-objective optimization strategy are adaptively adjusted to generate an updated multi-objective optimization strategy. In the next control cycle, the updated damage evolution model and the updated multi-objective optimization strategy are repeatedly executed to form a life extension closed-loop control system.

[0014] Optionally, in the seventh implementation of the first aspect of the present invention, the method further includes performing the following steps based on the remaining lifetime prediction results and health indicators: assessing the remaining safe operating window of key components according to a preset lifetime warning threshold and the remaining lifetime prediction results, and generating a component lifetime warning level; formulating differentiated operation mode plans for different warning levels based on the component lifetime warning level and the historical execution records of the collaborative control commands, and generating a tiered operation mode plan library; automatically switching the wind turbine's operation mode according to the tiered operation mode plan library when a change in the component lifetime warning level is detected, and generating and issuing a mode switching command; correcting the control parameters in the tiered operation mode plan library based on the operating effect after the mode switching command and the change in the damage rate of key components, and generating an optimized tiered operation mode plan library; and synchronously associating the optimized tiered operation mode plan library, the updated damage evolution model, and the updated multi-objective optimization strategy to form a lifetime extension collaborative management framework.

[0015] Optionally, in the eighth implementation of the first aspect of the present invention, after forming the life extension collaborative management framework, the following steps are performed: analyzing historical environmental data and unit operation data to identify typical severe operating condition combinations that accelerate the damage and degradation of key components, and generating a severe operating condition feature set; matching a conservative control strategy to each warning level in the graded operation mode contingency plan library based on the severe operating condition feature set, and generating a conservative strategy mapping table; monitoring current environmental parameters in real time and calculating the matching degree with the severe operating condition feature set; when the matching degree exceeds a preset threshold, triggering a temporary switch of the control strategy based on the conservative strategy mapping table. Generate condition-adaptive control commands; during the execution of the condition-adaptive control commands, synchronously record the actual cumulative damage data of key components and compare it with the predicted damage data of the damage evolution model under the corresponding operating conditions to generate a model prediction error assessment report; based on the model prediction error assessment report, reverse-calibrate the parameters related to typical severe operating conditions in the damage evolution model to generate a condition-adaptive damage evolution model; use the condition-adaptive damage evolution model to further optimize the control parameters in the graded operation mode contingency plan library, thereby updating the collaborative management framework and generating the final life extension management framework.

[0016] The mechanism of this invention is as follows: the entire method uses sensor measured data to calibrate the mechanism model online, realizing predictive maintenance that is deeply integrated with the long-term operating characteristics of the wind field; Beneficial effects: It can more accurately predict the remaining lifespan of key components, providing a scientific basis for the maintenance and control of wind turbine units. Compared with traditional static lifespan prediction methods, this dynamic model can adapt to the dynamic changes in the unit's operating status, improving the accuracy and reliability of predictions and avoiding over-maintenance or under-maintenance problems caused by inaccurate lifespan predictions; This effectively resolves the conflict between power generation efficiency and component lifespan in traditional control methods, achieving a balanced optimization of both. It can increase wind turbine power generation while reducing the rate of fatigue damage accumulation in key components, extending turbine lifespan, and improving the economy and reliability of wind power generation. This allows the wind turbine's control strategy to dynamically and adaptively adjust based on actual operating conditions, continuously improving the accuracy and effectiveness of control. The closed-loop control system can promptly detect and correct deviations in the control process, ensuring the unit is always in optimal operating condition and further extending its service life.

[0017] This enables wind turbines to automatically identify and adapt to harsh operating conditions, promptly adjusting control strategies to reduce damage to critical components and improve operational reliability and lifespan in complex environments. Adaptive control fully leverages the potential of wind turbines, enhancing power generation efficiency under varying operating conditions while ensuring safety. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an embodiment of the wind turbine adaptive life extension control method based on reinforcement learning in this invention. Figure 2 A schematic diagram of the adaptive life extension control process for wind turbine units; Figure 3 This is a schematic diagram of an embodiment of the wind turbine adaptive life extension control device based on reinforcement learning in this invention. Detailed Implementation

[0019] This invention provides an adaptive lifespan extension control method for wind turbines based on reinforcement learning, offering a reliable technical means to extend the lifespan of wind turbines. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0020] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figures 1-2One embodiment of the wind turbine adaptive life extension control method based on reinforcement learning in this invention includes: 101. Collect wind turbine operating status parameters and key component health status parameters to generate multi-dimensional status information containing real-time operating data and health assessment results; It is understood that the execution subject of this invention can be a wind turbine adaptive life extension control device based on reinforcement learning, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0021] Specifically, the wind speed, rotational speed, power, pitch angle and generator torque parameters of the wind turbine are acquired in real time during the operation of the wind turbine, and a real-time operation dataset is generated. Simultaneously collect vibration signals, temperature signals, and structural strain signals of key components to generate raw health monitoring datasets; Time-frequency domain analysis was performed on the vibration signals in the original health monitoring dataset to extract characteristic frequency components and vibration energy distribution, and vibration characteristic indicators were generated. Trend analysis and heat load assessment are performed on the temperature signals in the original health monitoring dataset, and the data are normalized in combination with the ambient temperature to generate temperature health indicators. Rainflow counting and load spectrum analysis were performed on the structural strain signals in the original health monitoring dataset to calculate the cumulative fatigue damage and generate health indicators. The real-time running dataset, vibration characteristic indicators, temperature health indicators, and health indicators are fused together to generate multi-dimensional state information to characterize the overall state of the unit.

[0022] 102. Based on multidimensional state information, establish a damage evolution model based on physical failure mechanism, and generate remaining life prediction results and health indicators of key components based on the model. Specifically, based on the material properties, geometric configuration, and design load spectrum of key components, a physical failure mechanism model describing their fatigue, wear, or crack propagation laws is constructed to generate a basic failure model; By utilizing real-time operational data and health indicators from multidimensional state information, the equivalent load and stress cycle borne by key components under current operating conditions are calculated, and a real-time load spectrum is generated. The real-time load spectrum is input into the basic failure model to calculate the damage increment caused by the current operating state, and then superimposed with the historical accumulated damage to generate an updated damage accumulation state. Based on the updated damage accumulation state and the failure threshold defined in the basic failure model, the running time or number of cycles required to reach the failure threshold under a preset future operating scenario is estimated, and a preliminary remaining lifetime prediction value is generated. The preliminary remaining life prediction is corrected by combining vibration characteristic indicators and temperature health indicators. The health index is generated by weighted fusion of the damage accumulation state and the corrected remaining life to quantify the overall degradation of the component.

[0023] 103. Based on the remaining lifetime prediction results and health indicators, design a multi-objective optimization strategy that includes power generation efficiency optimization and fatigue load control objectives; Specifically, based on the numerical range of the health index and the urgency of the remaining life prediction results, the priority relationship between the power generation efficiency optimization target and the fatigue load control target is dynamically determined, and dynamic weighting coefficients are generated. Based on wind speed and power data from real-time operation data, an objective function for optimizing power generation efficiency is established with the goal of maximizing power generation. Based on health indicators and real-time load spectrum, a fatigue load control objective function is established with the goal of minimizing the cumulative fatigue damage rate of key components. The objective function for power generation efficiency optimization and the objective function for fatigue load control are integrated, and the two are weighted and fused using dynamic weighting coefficients to generate a comprehensive objective function. Based on the design safety boundaries of key components and current operating conditions, operational safety constraints are set for pitch angle, generator torque, and stress of key components. The comprehensive objective function and the operational safety constraints are combined and encapsulated to generate a multi-objective optimization strategy with a clear optimization direction and boundary range for subsequent solution.

[0024] Furthermore, after designing the multi-objective optimization strategy, the following steps are performed: Based on the real-time load spectrum and the material SN curve or crack propagation rate curve of key components, the theoretical fatigue damage threshold of each component under the current operating conditions is calculated, and component-level load safety boundaries are generated. Based on health indicators and remaining life prediction results, the load safety boundary of components in the early warning state is dynamically adjusted to generate a dynamic load safety boundary that takes into account the actual health state of the components. The dynamic load safety boundary is used as a new hard constraint to supplement and replace some of the original operational safety constraints in the multi-objective optimization strategy, generating an updated multi-objective optimization strategy with integrated health state constraints. When using an adaptive search algorithm to solve the problem, priority is given to ensuring that the generated candidate control action sequences meet the dynamic load safety boundary, thereby guiding the search process toward a control region that is more friendly to vulnerable components and generating a health-oriented search space. The rationality of the dynamic load safety boundary is verified by using the feedback after the execution of the cooperative control command obtained in the health-oriented search space. Based on the verification results, it is fine-tuned to generate a dynamic load safety boundary library that has been calibrated in the field, which is used for the formulation of subsequent optimization strategies.

[0025] 104. An adaptive search algorithm is used to solve the multi-objective optimization strategy and generate coordinated control commands for coordinating pitch and torque actions; Specifically, based on the current multidimensional state information and multi-objective optimization strategy, a set of candidate control action sequences containing the combination of pitch angle and generator torque is initialized to generate an initial action population; Based on the operational safety constraints, the feasibility of each action sequence in the initial action population is checked, and sequences that do not meet the constraints are eliminated to generate a population of feasible actions. The performance of each action sequence in the action population is calculated using a comprehensive objective function to evaluate its overall performance in terms of power generation efficiency and fatigue load control, and to generate performance evaluation results. Based on the performance evaluation results, an adaptive mechanism based on sorting and selection is used to adjust the search direction and step size, and a new generation of action population is generated through crossover and mutation operations. Repeat the feasibility check, performance calculation and adaptive adjustment steps until the performance improvement of the action population is less than the preset threshold or the maximum number of iterations is reached, and generate an optimized action sequence. The optimal action combination with the best comprehensive objective function value is selected from the optimized action sequence. The corresponding pitch angle setpoint and generator torque setpoint are extracted to generate a coordinated control command for the current control cycle.

[0026] 105. Execute coordinated control commands and dynamically update the damage evolution model and optimization strategy based on the unit status feedback after command execution to form a life extension closed-loop control system.

[0027] Specifically, the coordinated control commands are sent to the pitch system and generator torque control system of the wind turbine and executed to generate corresponding actual control actions; After executing the coordinated control command, the wind turbine operating status parameters and key component health status parameters for a new monitoring cycle are collected to generate unit status feedback data for evaluating the control effect. Based on unit status feedback data, the actual impact of coordinated control commands on the fatigue damage accumulation rate of key components and power generation efficiency is analyzed, and a control effect evaluation report is generated. Based on the control effect evaluation report, determine whether the actual damage evolution trend of the key components is consistent with the predicted trend of the damage evolution model, and calibrate the relevant parameters in the damage evolution model according to the deviation to generate an updated damage evolution model. The health status of key components is reassessed using the updated damage evolution model, generating updated health indices and remaining life prediction results. Based on the updated health index and remaining life prediction results, the weight coefficients of the objective function or the operational safety constraints in the multi-objective optimization strategy are adaptively adjusted to generate an updated multi-objective optimization strategy. In the next control cycle, the updated damage evolution model and the updated multi-objective optimization strategy are adopted to repeatedly execute the steps of state monitoring, lifetime prediction, strategy solving and control command generation, thereby forming a lifetime-extending closed-loop control system that can continuously self-optimize based on execution feedback.

[0028] 106. Based on the remaining life expectancy prediction results and health indicators, perform the following steps: Based on the preset lifespan warning threshold and the remaining lifespan prediction results, the remaining safe operating window of key components is assessed, and the component lifespan warning level is generated. Based on the historical execution records of component lifespan warning levels and coordinated control commands, differentiated operation mode plans are formulated for different warning levels, and a hierarchical operation mode plan library is generated. When a change in the component life warning level is detected, the wind turbine's operating mode is automatically switched according to the graded operation mode contingency plan library, and a mode switching command is generated and issued. Based on the operational effects after the mode switching command and the changes in the damage rate of key components, the control parameters in the hierarchical operation mode contingency plan library are modified to generate an optimized hierarchical operation mode contingency plan library. The optimized hierarchical operation mode contingency plan library, the updated damage evolution model, and the updated multi-objective optimization strategy are synchronously linked to form a life extension collaborative management framework that integrates status early warning, adaptive adjustment of operation mode, and optimization of control strategy.

[0029] Furthermore, after establishing the lifespan extension collaborative management framework, the following steps are performed: By analyzing historical environmental data and unit operation data, typical severe operating condition combinations that accelerate the damage and degradation of key components are identified, and a severe operating condition feature set including parameters such as wind speed, turbulence intensity and temperature is generated. Based on the severe working condition feature set, one or more preset conservative control strategies for mitigating damage accumulation are matched in the graded operation mode contingency plan library for each warning level, and a conservative strategy mapping table is generated. Real-time monitoring of current environmental parameters and matching degree calculation with severe working condition feature set. When the matching degree exceeds the preset threshold, a temporary switch of control strategy based on conservative strategy mapping table is triggered, and working condition adaptive control command is generated. During the execution of adaptive control commands, the actual cumulative damage data of key components is recorded synchronously and compared with the predicted damage data of the damage evolution model under the corresponding working conditions to generate a model prediction error assessment report. Based on the model prediction error assessment report, the parameters related to typical severe working conditions in the damage evolution model are reverse-calibrated to generate a working condition adaptive damage evolution model. By utilizing a condition-adaptive damage evolution model, the control parameters in the graded operation mode contingency plan library are further optimized, thereby updating the collaborative management framework and generating a final life extension management framework with environmental condition adaptability.

[0030] In this embodiment of the invention, operational status and key component health parameters are collected to generate multi-dimensional status information from multiple dimensions. This comprehensively and accurately reflects the overall status of the unit, providing a solid and reliable data foundation for subsequent life prediction and control strategy formulation, avoiding the one-sidedness of single-parameter evaluation. A physical failure mechanism model is constructed based on the material properties and geometric configuration of key components. Real-time load spectra are generated by combining real-time data, thereby predicting remaining life and generating health indicators. Compared with traditional empirical models, this model more accurately reflects the actual damage evolution of components, improves prediction accuracy, and provides an accurate basis for rationally formulating control strategies. The priority of power generation efficiency optimization and fatigue load control objectives is dynamically determined based on the health indicators and remaining life prediction results. Dynamic weighting coefficients are generated, and the two objective functions are integrated and operational safety constraints are set to form a multi-objective optimization strategy. This strategy can flexibly adjust the optimization direction according to the actual state of the unit, balancing power generation efficiency and component fatigue load control while ensuring safe unit operation, thus achieving overall unit performance optimization. An adaptive search algorithm is employed to solve multi-objective optimization strategies. By initializing candidate control action sequences, and through feasibility verification, performance calculation, and adaptive adjustment, an optimized action sequence is generated. The optimal action combination is selected to generate coordinated control commands. This algorithm can adaptively adjust the search direction and step size based on the unit's status, quickly finding coordinated control commands that meet the multi-objective requirements, thus improving control efficiency and effectiveness. After executing the coordinated control commands, the damage evolution model and optimization strategy are dynamically updated based on unit status feedback, forming a lifespan extension closed-loop control system. This system can continuously self-optimize based on execution feedback, constantly adapting to changes in unit operating status, ensuring the control strategy remains effective, extending the service life of key unit components, and improving unit operational reliability and economy. Based on remaining lifespan prediction results and health indicators, the remaining safe operating window of components is assessed, generating component lifespan warning levels and establishing a tiered operating mode contingency plan library. When the warning level changes, the operating mode is automatically switched, and the control parameters in the contingency plan library are adjusted based on the operating effect, achieving differentiated and precise management for different warning levels and effectively ensuring the safe operation of the unit under different conditions. Historical data is analyzed to identify a set of severe operating conditions. A mapping table is generated by matching conservative control strategies with the tiered operation mode contingency plan library. Environmental parameters are monitored in real time to trigger temporary switching of control strategies and generate condition-adaptive control commands. During command execution, actual damage data is recorded to calibrate the damage evolution model. The tiered operation mode contingency plan library is then optimized to form a final life extension management framework with environmental condition adaptability. This enables the unit to maintain good operating condition under complex and changing environmental conditions, further extending its service life.

[0031] Figure 3 This is a schematic diagram of a reinforcement learning-based adaptive lifespan extension control device for wind turbines, provided in an embodiment of the present invention. The reinforcement learning-based adaptive lifespan extension control device 200 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 210 (e.g., one or more processors) and a memory 220, and one or more storage media 230 (e.g., one or more mass storage devices) for storing application programs 233 or data 232. The memory 220 and storage media 230 can be temporary or persistent storage. The program stored in the storage media 230 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the reinforcement learning-based adaptive lifespan extension control device 200. Furthermore, the processor 210 may be configured to communicate with the storage media 230 and execute the series of instruction operations in the storage media 230 on the reinforcement learning-based adaptive lifespan extension control device 200.

[0032] The reinforcement learning-based adaptive lifespan extension control device 200 for wind turbines may also include one or more power supplies 220, one or more wired or wireless network interfaces 250, one or more input / output interfaces 260, and / or one or more operating systems 231, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated structure of the wind turbine adaptive life extension control device based on reinforcement learning does not constitute a limitation on the wind turbine adaptive life extension control device based on reinforcement learning. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0033] The present invention also provides a wind turbine adaptive life extension control device based on reinforcement learning. The wind turbine adaptive life extension control device based on reinforcement learning includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the wind turbine adaptive life extension control method based on reinforcement learning in the above embodiments.

[0034] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the reinforcement learning-based wind turbine adaptive life extension control method.

[0035] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0036] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0037] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wind turbine adaptive lifespan extension control method based on reinforcement learning, characterized in that, include: Collect wind turbine operating status parameters and key component health status parameters to generate multi-dimensional status information, which includes real-time operating data and health assessment results. Based on the multidimensional state information, a damage evolution model is established, and based on the damage evolution model, the remaining life prediction results and health indicators of key components are generated. Based on the remaining lifetime prediction results and health indicators, a multi-objective optimization strategy is designed, including power generation efficiency optimization and fatigue load control objectives. After designing the multi-objective optimization strategy, the following steps are performed: Based on the real-time load spectrum and the material SN curve or crack propagation rate curve of the key components, the theoretical fatigue damage threshold of each component under the current operating conditions is calculated, and the component-level load safety boundary is generated. Based on the health index and remaining life prediction results, the load safety boundary of the component in the early warning state is dynamically adjusted to generate a dynamic load safety boundary. The dynamic load safety boundary is used as a new hard constraint to supplement and replace some of the original operational safety constraints in the multi-objective optimization strategy, thereby generating an updated multi-objective optimization strategy with integrated health state constraints. Solving the multi-objective optimization strategy generates coordinated control commands for coordinating pitch and torque actions. When using an adaptive search algorithm, priority is given to ensuring that the generated candidate control action sequences meet the dynamic load safety boundary, thereby guiding the search process toward a control region that is more friendly to vulnerable components and generating a health-oriented search space. The rationality of the dynamic load safety boundary is verified by using the feedback after the execution of the cooperative control command obtained by solving in the health-oriented search space, and fine-tuned based on the verification results to generate a dynamic load safety boundary library. The coordinated control commands are executed, and the damage evolution model and optimization strategy are dynamically updated based on the unit status feedback after the commands are executed, forming a life extension closed-loop control system.

2. The wind turbine adaptive lifespan extension control method based on reinforcement learning according to claim 1, characterized in that, include: Real-time acquisition of wind speed, rotational speed, power, pitch angle, and generator torque parameters during wind turbine operation, generating real-time operation datasets; Simultaneously collect vibration signals, temperature signals, and structural strain signals of key components to generate raw health monitoring datasets; Time-frequency domain analysis is performed on the vibration signals in the original health monitoring dataset to extract characteristic frequency components and vibration energy distribution, and to generate vibration characteristic indicators. Temperature health indicators are generated by performing trend analysis and heat load assessment on the temperature signals in the original health monitoring dataset. Rainflow counting and load spectrum analysis were performed on the structural strain signals in the original health monitoring dataset to calculate the cumulative fatigue damage and generate health indicators. The real-time running dataset, vibration characteristic indicators, temperature health indicators, and health indicators are fused together to generate multi-dimensional state information.

3. The wind turbine adaptive life extension control method based on reinforcement learning according to claim 2, characterized in that, include: Based on the material properties, geometric configuration, and design load spectrum of key components, a physical failure mechanism model is constructed to generate a basic failure model; Using the real-time operating data and health indicators in the multi-dimensional state information, the equivalent load and stress cycle borne by the key components under the current operating conditions are calculated, and a real-time load spectrum is generated. The real-time load spectrum is input into the basic failure model to calculate the damage increment caused by the current operating state, and then superimposed with the historical accumulated damage to generate an updated damage accumulation state. Based on the updated damage accumulation state and the failure threshold defined in the basic failure model, the running time or number of cycles required to reach the failure threshold under a preset future operating scenario is estimated, and a preliminary remaining lifetime prediction value is generated. The preliminary remaining life prediction value is corrected by combining the vibration characteristic index and the temperature health index, and a health index is generated by combining the damage accumulation state and the corrected remaining life.

4. The wind turbine adaptive life extension control method based on reinforcement learning according to claim 3, characterized in that, include: Based on the numerical range of the health index and the urgency of the remaining life prediction results, the priority relationship between the power generation efficiency optimization target and the fatigue load control target is dynamically determined, and dynamic weighting coefficients are generated. Based on the wind speed and power data in the real-time operating data, a power generation efficiency optimization objective function is established with the goal of maximizing power generation. Based on the aforementioned health indicators and real-time load spectrum, a fatigue load control objective function is established with the goal of minimizing the cumulative fatigue damage rate of key components. The power generation efficiency optimization objective function and the fatigue load control objective function are integrated, and the two are weighted and fused using the dynamic weighting coefficients to generate a comprehensive objective function; Based on the design safety boundaries of key components and current operating conditions, operational safety constraints are set for pitch angle, generator torque, and stress of key components. The comprehensive objective function is combined and encapsulated with the operational safety constraints to generate a multi-objective optimization strategy.

5. The wind turbine adaptive lifespan extension control method based on reinforcement learning according to claim 4, characterized in that, The process employs an adaptive search algorithm to solve the multi-objective optimization strategy, generating coordinated control commands for coordinating pitch and torque actions, including: Based on the current multidimensional state information and multi-objective optimization strategy, a set of candidate control action sequences containing the combination of pitch angle and generator torque is initialized to generate an initial action population. Based on the aforementioned operational safety constraints, a feasibility test is performed on each action sequence in the initial action population, sequences that do not meet the constraints are eliminated, and a feasible action population is generated. The performance of each action sequence in the actionable population is calculated using the comprehensive objective function to generate performance evaluation results. Based on the performance evaluation results, the search direction and step size are adjusted, and a new generation of action population is generated through crossover and mutation operations. Repeat the process until the performance improvement of the action population is less than a preset threshold or the maximum number of iterations is reached, then generate an optimized action sequence. The optimal action combination with the best comprehensive objective function value is selected from the optimized action sequence, and its corresponding pitch angle setpoint and generator torque setpoint are extracted to generate a coordinated control command for the current control cycle.

6. The wind turbine adaptive life extension control method based on reinforcement learning according to claim 5, characterized in that, The process of executing the coordinated control commands and dynamically updating the damage evolution model and optimization strategy based on the unit status feedback after command execution to form a life extension closed-loop control system includes: The coordinated control command is sent to the pitch system and generator torque control system of the wind turbine and executed to generate corresponding actual control actions. After executing the collaborative control command, the wind turbine operating status parameters and key component health status parameters for a new monitoring cycle are collected to generate turbine status feedback data. Based on the unit status feedback data, the actual impact of the coordinated control commands on the fatigue damage accumulation rate of key components and power generation efficiency is analyzed, and a control effect evaluation report is generated. Based on the control effect evaluation report, determine whether the actual damage evolution trend of the key components is consistent with the predicted trend of the damage evolution model, and calibrate the relevant parameters in the damage evolution model according to the deviation to generate an updated damage evolution model. The health status of key components is reassessed using the updated damage evolution model, generating updated health indices and remaining life prediction results. Based on the updated health index and remaining life prediction results, the weight coefficients of the objective function or the operational safety constraints in the multi-objective optimization strategy are adaptively adjusted to generate an updated multi-objective optimization strategy. In the next control cycle, the updated damage evolution model and the updated multi-objective optimization strategy are adopted and repeatedly executed to form a life-extending closed-loop control system.

7. The wind turbine adaptive life extension control method based on reinforcement learning according to claim 1, characterized in that, It also includes performing the following steps based on the remaining life expectancy prediction results and health indicators: Based on the preset lifespan warning threshold and the remaining lifespan prediction results, the remaining safe operating window of key components is assessed, and a component lifespan warning level is generated. Based on the component lifespan warning level and the historical execution records of the coordinated control commands, differentiated operation mode plans are formulated for different warning levels, and a hierarchical operation mode plan library is generated. When a change in the lifespan warning level of the component is detected, the wind turbine's operating mode is automatically switched according to the graded operation mode contingency plan library, and a mode switching command is generated and issued. Based on the operational effects after the mode switching command and the change in the damage rate of key components, the control parameters in the hierarchical operation mode contingency plan library are modified to generate an optimized hierarchical operation mode contingency plan library. The optimized hierarchical operation mode contingency plan library, the updated damage evolution model, and the updated multi-objective optimization strategy are synchronously linked to form a collaborative management framework for life extension.

8. The wind turbine adaptive life extension control method based on reinforcement learning according to claim 7, characterized in that, After establishing the aforementioned lifespan extension collaborative management framework, the following steps are performed: By analyzing historical environmental data and unit operation data, typical severe operating condition combinations that accelerate the damage and degradation of key components are identified, and a severe operating condition feature set is generated. Based on the severe working condition feature set, a conservative control strategy is matched for each early warning level in the graded operation mode contingency plan library, and a conservative strategy mapping table is generated. Real-time monitoring of current environmental parameters and matching degree calculation with the harsh working condition feature set; when the matching degree exceeds a preset threshold, triggering a temporary switch of control strategy based on the conservative strategy mapping table and generating working condition adaptive control instructions. During the execution of the working condition adaptive control command, the actual damage accumulation data of key components is recorded synchronously and compared with the predicted damage data of the damage evolution model under the corresponding working condition to generate a model prediction error evaluation report. Based on the model prediction error assessment report, the parameters related to typical severe working conditions in the damage evolution model are reverse-calibrated to generate a working condition adaptive damage evolution model. By utilizing the condition-adaptive damage evolution model, the control parameters in the graded operation mode contingency plan library are further optimized, thereby updating the collaborative management framework and generating the final life extension management framework.

Images