Research And Implementation On Data-Driven Optimization Methods For Wind Turbine Parameters

Wind energy is currently one of the most successful forms of renewable energy for commercialization.Wind turbines are the core of the entire wind energy industry.At present,one of the major problems in the wind energy industry is that wind turbines have low efficiency in converting wind energy into electrical energy,because of the complex real-world conditions and the immaturity of control technologies.Therefore,under the huge installed capacity,optimization of the power output of wind turbines will bring considerable benefits.Traditional wind turbine control methods rely on human experience and preset system control curve,so it's difficult for them to deal with complex and variable turbine operating conditions.At present,most wind farms are equipped with a Supervisory Control and Data Acquisition(SCADA)system.In the long-term operation,a large number of actual operational data of wind turbines have been accumulated.Novel technologies for wind turbine modeling and parameters optimization problem has emerged.These technologies are based on data mining theories,they utilize the actual wind turbine operational data accumulated in the wind farm,and excavate the nonlinear relationship between the input variables and the output variables.Input variables are turbine working environment parameters,such as wind speed and control parameters like pitch angle and generation torques.Output variables are turbine performance parameters,such as output power and rotor speed.Then,optimization algorithms are used to obtain optimized control parameters for specific working conditions.Therefore,based on operational data of a wind turbine,this thesis' s work includes two parts.Firstly,this thesis tries to model and compute optimized parameters for wind turbine at a single time step.Secondly,there is no research on wind turbine modeling and parameters optimization over temporal segments.So from a novel perspective,this thesis discusses modeling and parameters optimization methods for wind turbines over temporal segments.This thesis mainly focuses on the following contents:1.For the parameters optimization at a single time step,a preprocessing method for wind turbine operational data are proposed firstly.The preprocessing method effectively removes the noise in the raw data.Then based on domain knowledge,suitable attribute variables are selected for modeling the wind turbine power generation system.A modified recurrent neural network(RNN)—long short-term memory(LSTM)network,is used to learn the dynamic model of the wind turbine's power output process.Experiments show that the model has a higher accuracy compared to baselines.Finally,under the optimization goal of maximizing power efficiency and stability,an improved differential evolution(DE)algorithm is proposed.The algorithm is applied to the dynamic model to obtain optimized parameters.Experiments show that the improved DE algorithm is faster than baselines and the utilization the computed optimized parameters can improve the performance of the wind turbine.2.For the parameters optimization over temporal segments,this thesis firstly uses a deep generation model variational auto-encoder(VAE)to learn the dynamic identification model of the wind turbine.The encoder learns the distribution over latent codes Z conditioned on input variables and system output variables in the past segment and future segment.The decoder learns to reconstruct system output variables in the future segment from the past input variables,output variables,future input variables and a sample from Z.After training is complete,the decoder will allow us to predict system output variables in the future segment using past input variables,output variables,future input variables,as desired,sampling latent codes form the distribution Z.Furthermore,a modified DE algorithm that is suitable for segment model is presented.Experimental results reveal that the segment model is more accurate than chained one-step models.And the computed optimized parameters are helpful for improvement of wind turbine performance over temporal segments.