Data-driven Performance Improvement And Fault Early Warning For Large Scale Wind Turbine Systems

With the rapid development of the economy and human society,the ever-accelerating depletion of energy resources and increasingly serious environmental pollution have always become urgent problems to be faced and tackled for a long time.Therefore,the development of environmentally friendly renewable clean energy has been highly valued and widely concerned by the international community.As a new type of renewable clean energy with the most mature industrialization conditions,wind energy has been widely recognized and favored all over the world due to its characteristics of less pollution and rich reserves.Wind power generation technologies have also made rapid progress in a large number of state-of-the-art relevant researches in both academic literature and industrial applications in recent years.With the gradual large-scale and more intelligent development of wind turbines,the low generation performances of wind turbines in complex operation environments,the high operation and maintenance expenditure caused by remote deployment wind farms,and the high risks faced by key components of wind turbines make the economic operation of wind farm increasingly prominent.Therefore,for large horizontal-axis variable-speed wind turbines,researches on data analysis technology of wind turbines,including quantitative analysis of power generation performance,optimization and improvement of power generation performance,accurate early warning of frequent faults,play positive roles in promoting the intelligent operation and maintenance of wind power industry.With the global promotion of wind power digitization,the abundant data resources in the supervisory control and data acquisition(SCADA)system of a wind turbine have brought fresh vigor and vitality into the rapid iteration and innovation of wind turbine data analysis technologies,and also provide an unlimited possibility for realizing the stable,reliable,safe and efficient operation of the wind turbine from the perspective of data-driving and data-mining.Combined with the state-of-the-art literature in various fields,this thesis puts forward a series of solutions to the problems of data quality improvement,power generation performance evaluation,power generation performance improvement,frequent fault early warning and so on.The effectiveness,accuracy,reliability,and enforceability of various algorithms introduced in those solutions are analyzed,verified,and testified based on both simulation operation data set and real operation data set of a certain type of wind turbine.The research works of this thesis mainly can be summarized as follows:1.The background and development trend of wind power generation are first provided,and the development status and related works of wind turbine data analysis technologies are also introduced in details.2.Research on the preprocessing of wind turbine SCADA data.Firstly,the spatial and temporal distribution characteristics of wind turbine SCADA data are studied for the purpose of analyzing temporal and spatial correlations of data variables.Secondly,in order to identify abnormal operation states of spatial scattered data of wind turbine power curve,a wind turbine preset performance-based anomaly detection algorithm and a distance-density mixed information-based outlier detection algorithm are proposed respectively for improving the reliability of wind turbine SCADA data.Furthermore,a DAE and bidirectional gated recurrent unit(Bi-GRU)based missing data imputation algorithm is proposed to achieve the highest missing data reconstruction accuracy on the basis of meeting the continuity requirements of time series SCADA data.Finally,the effectiveness and accuracy of the two proposed data preprocessing algorithms are verified on both simulation data set and real data set of a certain type of wind turbine.3.Research on the generation performance analysis based on wind turbine SCADA data.Firstly,based on the preprocessed SCADA data of wind turbines,a wind turbine power curve modeling algorithm is designed based on the least square B-spline approximation technology.The high reliable power curve model of wind turbines can be established by introducing steps including wind speed interval partitioning,centroids calculation and dominant point selection.Secondly,an iterative modeling procedure of power curve for standardized correction of wind turbulence intensity is performed for the purpose of correlation decoupling between active power output and wind turbulence intensity,thus improving the accuracy and authenticity of power curve modeling results.Moreover,the relationships between the values of hyperparameters and the modeling performance are analyzed,and an optimal selection strategy of hyperparameters is formulated to maximize power curve modeling performance.Finally,the proposed power curve modeling algorithm is compared with two commonly-used power curve modeling methods in academia and industry,and the effectiveness and accuracy of the modeled power curve are testified based on the real data set of a certain type of wind turbine.4.Research on the identification of yaw error misalignment of the wind turbine.Firstly,the concept and causes of yaw error misalignment are introduced in details,and the correlation between yaw misalignment and the power generation performance of wind turbine is also analyzed quantitatively.Secondly,from the perspective of data mining,two online identification algorithms of yaw error misalignment are designed for low-frequency sampling SCADA data and high-frequency sampling SCADA,respectively.For low-frequency SCADA data,power curves under different yaw error intervals are modeled,and a power curve quantitative analysis approach is further implemented to evaluate the power generation performance under all yaw error intervals,thus the range of yaw error misalignment can be easily estimated.For high-frequency SCADA data,small range partitioning based on both wind speed and rotor speed for SCADA data in low wind speed section is applied,and the correlation function between measured yaw error and active power output is fitted based on least-square estimation technology,and the accurate identification result of yaw error misalignment under a certain degree of confidence can be determined.Finally,the effectiveness,accuracy and enforceability of the two proposed algorithms are verified based on both simulation data set and real data set of a certain type of wind turbine.5.Research on the early warning on key components over-temperature fault of the wind turbine.Firstly,the causes and manifestations of wind turbine temperature fault are summarized.Secondly,according to the early warning requirements on key components over-temperature faults of the wind turbine,a data-driven early warning algorithm combining normal behavior model,residual control chart and early warning discrimination strategy is proposed.The normal behavior model is mainly utilized to extract the correlation between temperature variables closely related to the over-temperature fault and other variables under wind turbine normal operation.With the normal behavior model residual sequence defined as the difference between the real value and the estimated value of the normal behavior model,the residual control chart is adopted for detecting abnormal changes on it.The early warning discrimination strategy is used to further improve the robustness and reliability of identified results of abnormal changes on the residual sequence.In addition,the normal behavior models constructed by different machine learning technologies are combined with different residual control charts,and the optimal combination is selected as the final structure of the fault early warning algorithm by comprehensively considering the normal behavior modeling performance and residual control chart detection performance,thus the performance of the fault early warning algorithm can be maximized.Finally,the early warning application results on two key components over-temperature faults are validated on real SCADA data of a wind farm in China.6.At the end of this thesis,the aforementioned research works are concluded,and some future research directions are prospected.