CN-121993347-A - Control method of wind farm pitch system

CN121993347ACN 121993347 ACN121993347 ACN 121993347ACN-121993347-A

Abstract

The invention relates to the technical field of wind power generation, in particular to a control method of a wind farm pitch system. The method comprises the steps of constructing a general strategy model based on meta learning, obtaining initial parameters irrelevant to fans through cross-fan meta training, supporting rapid fine adjustment adaptation, deploying intelligent agents taking the initial parameters as starting points on each fan, performing local strategy optimization by utilizing deep reinforcement learning, designing a composite reward function integrating power generation capacity, mechanical load and power smoothness, guiding the intelligent agents to achieve multi-objective dynamic balance, constructing a federal learning cooperative network, uploading model parameter update through encryption of each fan, and achieving full-field cooperative optimization through aggregation and distribution of a central server, and enabling the system to continuously perform self-adaptive optimization through closed loop cooperation of the technology. According to the invention, through fusion element learning, deep reinforcement learning and federal learning, the overall power generation efficiency, equipment reliability and operation intelligent level of the wind farm are remarkably improved.

Inventors

LI CHUNFENG
JIANG HUI
WU HUIJIE
Zhao Dongle
GAO YAQIANG
QI HONGBIN

Assignees

水发丰电能源(济南)有限公司

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (10)

1. The control method of the wind farm pitch system is characterized by comprising the following steps of: S100, constructing a general strategy model based on meta-learning, and performing cross-fan meta-training on heterogeneous fan historical data to enable the model to obtain initial parameters irrelevant to fans, wherein when a new fan is connected or working conditions are changed, the model can be subjected to quick fine adjustment adaptation only by collecting short-term operation data; s200, locally deploying an agent control unit on each fan, taking initial parameters of the general strategy model as a starting point, operating as edge computing nodes, continuously and interactively exploring in a simulation environment and a real-time system by using a deep reinforcement learning algorithm, and optimizing a control strategy through local data; S300, designing a multidimensional composite rewarding function, dynamically evaluating a variable pitch control effect, comprehensively considering multiple target indexes such as generating capacity, mechanical load, power output smoothness and the like by the function, and guiding an intelligent body to autonomously balance optimization priorities under different working conditions.
2. The control method of a wind farm pitch system according to claim 1, wherein the step 100 of meta-training adopts a model independent meta-learning framework, and the model initial parameters are optimized to enable the model initial parameters to quickly converge on various virtual working condition tasks, and the generalization capability of the dynamic characteristics of fresh air is enhanced through a gradient update mechanism.
3. The method for controlling a pitch system of a wind farm according to claim 2, wherein a working condition classifier is introduced in the meta-training process, and the current working condition is classified according to the real-time wind speed, turbulence intensity and load signal, and the closest initial parameter subset in the meta-model is dynamically selected to accelerate the fine tuning adaptation process.
4. The method for controlling a pitch system of a wind farm according to claim 1, wherein the deep reinforcement learning in S200 employs an actor-commentator architecture, wherein an actor network generates pitch motions, and the commentator network evaluates the motion values and stores interactive data through an experience playback buffer to improve learning stability.
5. The method for controlling a pitch system of a wind farm according to claim 4, wherein the actor network and the critics network are both in a deep neural network structure, and the network inputs comprise wind speed time sequence data, fan rotation speed, pitch angle and load measurement values, and outputs continuous pitch angle control instructions.
6. The method according to claim 1, wherein the specific form of the multidimensional composite reward function in S300 is a weighted sum function, wherein the generated energy reward is positively correlated with the actual power output, the mechanical load penalty is negatively correlated with the tower vibration amplitude and the blade stress, the power smoothness reward is negatively correlated with the power change rate, and the weight coefficient is dynamically adjustable according to the power grid scheduling requirement.
7. The method of controlling a wind farm pitch system according to claim 1, further comprising: S400, constructing a wind farm level federal learning cooperative network, wherein each fan intelligent agent encrypts model parameter update after local training, and distributes the model parameter update after global update is aggregated through a central server so as to enable a full-farm fan to implicitly learn wake interference rules and cooperative strategies; In the step S400, the federal learning cooperative network adopts an asynchronous aggregation strategy, so that partial fans are allowed to suspend parameter uploading due to communication delay, and a central server fuses available updating through a weighted average algorithm, so that the robustness and consistency of full-field model updating are ensured.
8. The method for controlling a pitch system of a wind farm according to claim 7, wherein differential privacy noise is added to parameter update in the federal learning process to protect local data privacy of each fan and prevent model degradation caused by malicious attack through a model version management mechanism.
9. The control method of a wind farm pitch system according to claim 1, wherein the agent control unit integrates a fault tolerance module, automatically switches to a backup control strategy based on historical data when a sensor fault or a communication interruption is detected, and triggers a model re-initialization process.
10. The method of controlling a wind farm pitch system according to claim 1, further comprising: s500, through closed loop cooperation of meta learning, deep reinforcement learning and federal learning, the variable pitch system realizes intelligent transition from perceived environment change to self-adaptive decision making, and dynamic balance of power generation efficiency, equipment service life and power grid requirements is continuously optimized.

Description

Control method of wind farm pitch system Technical Field The invention relates to the technical field, in particular to a control method of a wind farm pitch system. Background Wind energy is used as an important component of clean energy, and the efficient and stable utilization of the wind energy depends on the fine control of a wind farm. The pitch system is used as a key actuating mechanism of the wind generating set and is responsible for controlling the power captured by the fan and the born load by adjusting the pitch angle of the blades, and the control performance of the pitch system is directly related to the power generation efficiency, the equipment service life and the power grid stability. Traditional wind farm pitch control methods rely primarily on controllers designed based on fixed mathematical models, such as proportional-integral-derivative controllers and their modifications. However, in actual operation, wind resources have strong nonlinearities, time-variations and uncertainties, including complex turbulence, wind shear effects, and complex wake disturbances between units within a wind farm. This makes it difficult for conventional controllers based on fixed models and parameters to maintain optimal performance over a full range of operating conditions, often resulting in insufficient adaptability and reduced control accuracy in dynamically changing environments. In addition, the conventional control method generally uses a single target (such as maximum power point tracking) as a core, and it is difficult to effectively cooperatively process a plurality of conflicting control targets such as maximizing the power generation efficiency, minimizing the mechanical load, and smoothing the power output. In recent years, although research attempts are made to introduce a single-machine intelligent control algorithm, the model generalization capability is weak, the model is difficult to directly migrate to fans of different models or in different wind conditions, and a cooperative mechanism of a wind power plant level is lacking, so that the total output of the whole farm cannot be optimized from the system level, and adverse effects of wake flow are relieved. Therefore, a novel variable pitch control method which can adapt to dynamic environment, intelligently balance multi-target conflict and has field-level collaborative optimization capability is urgently needed to break through the technical bottleneck faced by the current wind power field operation. Thus, the prior art is still to be further developed. Disclosure of Invention The invention aims to overcome the technical defects and provide a control method of a wind farm pitch system so as to solve the problems in the prior art. In order to achieve the technical purpose, the invention provides a control method of a wind farm pitch system, which comprises the following steps: S100, constructing a general strategy model based on meta-learning, and performing cross-fan meta-training on heterogeneous fan historical data to enable the model to obtain initial parameters irrelevant to fans, wherein when a new fan is connected or working conditions are changed, the model can be subjected to quick fine adjustment adaptation only by collecting short-term operation data; s200, locally deploying an agent control unit on each fan, taking initial parameters of the general strategy model as a starting point, operating as edge computing nodes, continuously and interactively exploring in a simulation environment and a real-time system by using a deep reinforcement learning algorithm, and optimizing a control strategy through local data; S300, designing a multidimensional composite rewarding function, dynamically evaluating a variable pitch control effect, comprehensively considering multiple target indexes such as generating capacity, mechanical load, power output smoothness and the like by the function, and guiding an intelligent body to autonomously balance optimization priorities under different working conditions. Specifically, the meta-training in S100 adopts a model independent meta-learning framework, and makes the model initial parameters converge on various virtual working condition tasks rapidly by optimizing the model initial parameters, and enhances the generalization capability of the dynamic characteristics of the fresh air through a gradient update mechanism. Specifically, a working condition classifier is introduced in the meta-training process, the current working condition is classified according to the real-time wind speed, the turbulence intensity and the load signal, and the closest initial parameter subset in the meta-model is dynamically selected so as to accelerate the fine tuning adaptation process. Specifically, in the step S200, the deep reinforcement learning adopts an actor-critter architecture, wherein an actor network generates a pitching motion, and the critter network evaluates the motion value and stores interactive da