CN-122021244-A - Rapid modeling method and system for wind turbine generator based on data driving

CN122021244ACN 122021244 ACN122021244 ACN 122021244ACN-122021244-A

Abstract

A rapid modeling method and system for wind turbine generator based on data driving collect the operation data of SCADA system of wind turbine generator and preprocess to form time sequence data set, adopt the feature selection method of mutual information and random forest fusion to set up dual screening threshold value and keep key feature variable, build fan dynamic property agent model based on gate control circulation unit, train by Huber loss function, introduce physical constraint such as power-wind speed monotonicity, rotation speed-power property, pitch angle-power relation and so on, and fuse into loss function by penalty term, update model parameter in real time by recursive least square algorithm with forgetting factor, build standardized interface service by gRPC frame to realize integration with wind field control system, build real time residual monitoring, rolling cross verification and month model audit three-stage verification system. The method realizes high-precision, self-adaptive and deployable rapid modeling of the wind turbine generator, and remarkably shortens the modeling period.

Inventors

CHEN JINGUI
JIANG GUANGQIU
WANG JIANBO
CHEN ZHIXIONG
CAI ZHENGWEI
LIU XIBIN
WANG HAORAN
LUO ZHAN
CAI ZHAOBING

Assignees

长江三峡集团福建能源投资有限公司
福清海峡发电有限公司
三峡智控科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251223

Claims (16)

1. A rapid modeling method of a wind turbine generator based on data driving is characterized by comprising the following steps: s1, collecting operation data of a SCADA system of a wind turbine, and preprocessing the operation data to form a time sequence data set; s2, calculating the correlation between the operation parameters in the operation data and the core state variables of the wind turbine by adopting a feature selection method of fusion of the mutual information and the random forest, setting a dual screening threshold value, and reserving key feature variables; S3, constructing a fan dynamic characteristic proxy model based on the gating circulating unit, wherein the fan dynamic characteristic proxy model comprises a plurality of layers of gating circulating units and a layer of fully-connected output layer, and training by adopting a Huber loss function; S4, updating parameters of the fan dynamic characteristic proxy model in real time by adopting a recursive least square algorithm with forgetting factors, wherein the forgetting factors are initially set to be values in a preset range, and when the prediction error of the fan dynamic characteristic proxy model exceeds a set threshold value in a plurality of continuous periods, the forgetting factors are reduced; s5, introducing physical constraint in the training process of the fan dynamic characteristic proxy model, and adding a Huber loss function in a punishment term form; S6, constructing a standardized interface service through a gRPC framework, wherein the standardized interface service provides the functions of prediction, parameter updating and state query of the fan dynamic characteristic proxy model, and realizes the integration of the fan dynamic characteristic proxy model and a wind field control system; And S7, establishing a three-level verification system comprising real-time residual error monitoring, rolling cross verification and monthly model auditing, and starting an incremental retraining process to update the fan dynamic characteristic proxy model when the performance of the fan dynamic characteristic proxy model is continuously reduced.
2. The rapid modeling method of a wind turbine generator set based on data driving according to claim 1, wherein the dual screening threshold in step S2 includes a mutual information value threshold and a random forest feature importance score threshold, and the number of key feature variables is 15 to 20.
3. The rapid modeling method of a wind turbine generator set based on data driving according to claim 1, wherein in step S3, a wind turbine dynamic characteristic proxy model is trained by using AdamW optimizers, and Dropout layers are arranged between a gating circulation unit and a full connection layer to prevent overfitting.
4. The rapid modeling method of a wind turbine generator based on data driving according to claim 2, wherein the mutual information value threshold is 0.2, and the random forest feature importance score threshold is 0.05.
5. The rapid modeling method of a wind turbine generator set based on data driving according to claim 1, wherein in the step S4, when parameters of a fan dynamic characteristic proxy model are updated by a recursive least square algorithm with forgetting factors, an upper limit value of a parameter updating range is set, and the fan dynamic characteristic proxy model is prevented from being output unstably due to the fact that a single parameter adjustment amount exceeds a set range.
6. The method for rapid modeling of a wind turbine generator based on data driving of claim 1, wherein the physical constraints in step S5 include a power-wind speed monotonicity constraint, a rotational speed-power characteristic constraint, and a pitch angle-power relationship constraint.
7. The rapid modeling method of the wind turbine based on data driving according to claim 6 is characterized in that monotonically constant output electric power of the wind turbine under the same working condition along with increasing wind speed is required by power-wind speed monotonicity constraint, power value output by a dynamic characteristic proxy model of a fan is required by rotation speed-power characteristic constraint to fall within a reasonable range of a characteristic curve of the wind turbine, and output electric power of the wind turbine is required to be reduced along with increasing pitch angle in a region above rated wind speed by pitch angle-power relation constraint.
8. The rapid modeling method of the wind turbine generator based on data driving is characterized in that a self-adaptive weight adjustment strategy is adopted for punishment items of physical constraints in the step S5, the system monitors the meeting condition of each physical constraint in real time, when any physical constraint is violated, the weight coefficient of the corresponding punishment item is increased, the system dynamically adjusts the weight of the physical constraint according to the operation condition of the wind turbine generator, and the weight of the physical constraint is reduced in the process of starting, stopping or changing the operation condition of the wind turbine generator.
9. The rapid modeling method of a wind turbine generator system based on data driving according to claim 1, wherein in step S6, a standard interface service adopts a double-buffer data access architecture, a real-time data channel accesses a memory mapping area of a SCADA system of the wind turbine generator system through a shared memory, a batch processing data channel is connected with a real-time database through ODBC, and a lock-free buffer area is adopted between the two channels to realize data synchronization.
10. The method for rapid modeling of a wind turbine generator system based on data driving of claim 9, wherein the data acquisition delay of the real-time data channel is less than 10 milliseconds.
11. The rapid modeling method of the wind turbine generator set based on data driving according to claim 1 is characterized in that in step S7, prediction errors of a fan dynamic characteristic agent model are calculated according to fixed time granularity through real-time residual error monitoring, early warning is sent out when the prediction errors of a plurality of continuous monitoring periods exceed a set threshold value, rolling cross verification is conducted, the generalization capability of the fan dynamic characteristic agent model is evaluated through a sliding time window, and month model audit is conducted to comprehensively evaluate long-term performance of the fan dynamic characteristic agent model.
12. The rapid modeling method of the wind turbine generator set based on data driving according to claim 11, wherein the time granularity of real-time residual error monitoring is 1 minute, early warning is sent when the prediction error of 3 continuous monitoring periods exceeds a set threshold value, and rolling cross verification adopts a 24-hour sliding time window.
13. The rapid modeling method of a wind turbine generator set based on data driving according to claim 1, wherein in the step S7, the three-level verification system adopts an intelligent threshold adjustment strategy, the system adjusts an alarm threshold of a prediction error according to the turbulence intensity of a wind field, the error threshold is increased when the turbulence intensity of the wind field is increased, and the error threshold is reduced when the turbulence intensity of the wind field is reduced.
14. The rapid modeling method of a wind turbine generator set based on data driving according to claim 1, wherein when the incremental retraining process updates the fan dynamic characteristic proxy model parameters, important samples in historical data are reserved, and a memory playback mechanism is adopted to prevent the fan dynamic characteristic proxy model from forgetting historical knowledge.
15. The rapid modeling method of the wind turbine generator set based on data driving according to claim 1, wherein a fan dynamic characteristic proxy model is deployed on a Kubernetes container arrangement platform, and automatic capacity expansion and fault recovery of standardized interface service are achieved through a container mode.
16. A rapid modeling system for a wind turbine based on data driving, wherein a rapid modeling method for a wind turbine based on data driving according to any one of claims 1 to 15 is adopted, the system comprising: The data preprocessing module is used for collecting the operation data of the SCADA system of the wind turbine, preprocessing the operation data of the SCADA system of the wind turbine and generating a high-quality time sequence data set; the feature selection module is used for calculating the correlation between each operation parameter in the operation data of the SCADA system of the wind turbine and the core state variable of the wind turbine by adopting a feature selection method of fusion of the mutual information and the random forest, and outputting key feature variables according to the double screening conditions of the mutual information value threshold and the random forest feature importance score threshold; the model construction module is used for constructing a fan dynamic characteristic proxy model based on the gating circulating unit, wherein the fan dynamic characteristic proxy model comprises a plurality of layers of gating circulating units and a layer of fully-connected output layer, and training is carried out by adopting a Huber loss function; The online learning module is used for updating parameters of the fan dynamic characteristic proxy model in real time by adopting a recursive least square algorithm with forgetting factors, wherein the forgetting factors are initially set to be values in a preset range, and when the prediction error of the fan dynamic characteristic proxy model exceeds a set threshold value in a plurality of continuous periods, the forgetting factors are reduced; the physical constraint embedding module is used for introducing physical constraint in the training process of the fan dynamic characteristic proxy model and adding a Huber loss function in a punishment term form; the interface service module is used for constructing standardized interface service through gRPC frames, and the standardized interface service provides the functions of prediction, parameter updating and state query of the fan dynamic characteristic proxy model; The verification and check module is used for executing real-time residual error monitoring, rolling cross verification and monthly model audit, and when the performance of the fan dynamic characteristic proxy model is continuously reduced, an incremental retraining process is started to update the fan dynamic characteristic proxy model; all modules are deployed on a Kubernetes container arrangement platform through a micro-service architecture, so that high availability and elastic expansion of the system are realized.

Description

Rapid modeling method and system for wind turbine generator based on data driving Technical Field The invention belongs to the technical field of wind turbine modeling, and particularly relates to a rapid modeling method and system for a wind turbine based on data driving. Background At present, modeling of a wind turbine mainly depends on a white box or gray box model based on a physical principle, such as building a high-precision simulation model by using professional software such as Bladed, FAST and the like. The method is generally used in the design stage of the wind turbine generator, a mathematical model is constructed according to aerodynamics, structural mechanics and control theory, and the performance of the mathematical model is verified through simulation. However, the prior art has the following defects that firstly, the traditional method has a long modeling period, the traditional method relies on detailed physical parameters and complex simulation, the modeling process is time-consuming, and the requirements of engineering real-time control and quick operation and maintenance are difficult to meet. Secondly, the working condition adaptability is weak, the classical model parameters are fixed, and the adaptive adjustment is difficult to carry out along with the equipment aging, environmental change and other dynamic working conditions, so that the prediction accuracy is reduced along with the running time. Finally, the traditional model is very complex to deploy, the high-fidelity model often depends on specific software and hardware environments, seamless integration with the active wind turbine generator control system and the data platform is difficult to realize, and the deployment cost is high and the flexibility is poor. In summary, the technical problems to be solved by the invention are as follows: 1. how to utilize the actual operation data of the wind turbine generator to quickly construct a high-precision digital model, so that the modeling period is shortened; 2. How to realize the dynamic self-correction of model parameters, and adapt to variable working conditions such as aging of a unit, environmental change and the like; 3. How to design standardized interfaces and verification mechanisms, and support rapid engineering deployment and application of models in existing control systems. Disclosure of Invention The invention aims to solve the technical problem of providing a rapid modeling method and a rapid modeling system for a wind turbine based on data driving, which are used for realizing high-precision and self-adaptive digital modeling of pneumatic, transmission and control characteristics of the wind turbine by fusing multisource operation data with a lightweight machine learning algorithm. In order to solve the technical problems, the invention adopts the following technical scheme: A rapid modeling method of a wind turbine based on data driving comprises the following steps: S1, collecting operation data of a SCADA system of a wind turbine, and preprocessing the operation data to form a high-quality time sequence data set; S2, calculating the correlation between each operation parameter in the operation data and the core state variable of the wind turbine by adopting a feature selection method of fusion of the mutual information and the random forest, setting a dual screening threshold value, and reserving key feature variables; S3, constructing a fan dynamic characteristic proxy model based on the gating circulating unit, wherein the fan dynamic characteristic proxy model comprises a plurality of layers of gating circulating units and a layer of fully-connected output layer, and training by adopting a Huber loss function; S4, updating parameters of the fan dynamic characteristic proxy model in real time by adopting a recursive least square algorithm with forgetting factors, wherein the forgetting factors are initially set to be values in a preset range, and when the prediction error of the fan dynamic characteristic proxy model exceeds a set threshold value in a plurality of continuous periods, the forgetting factors are reduced; s5, introducing physical constraint in the training process of the fan dynamic characteristic proxy model, and adding a Huber loss function in a punishment term form; S6, constructing a standardized interface service through a gRPC framework, wherein the standardized interface service provides the functions of prediction, parameter updating and state query of the fan dynamic characteristic proxy model, and realizes the integration of the fan dynamic characteristic proxy model and a wind field control system; And S7, establishing a three-level verification system comprising real-time residual error monitoring, rolling cross verification and monthly model auditing, and starting an incremental retraining process to update the fan dynamic characteristic proxy model when the performance of the fan dynamic characteristic proxy model is continuousl