基于TL-DNN的风力机高保真尾流代理建模  

High-fidelity Wind Turbine Wake Surrogate Modeling Based on TL-DNN

作　　者：谭郡瑶 王强[1,2] 罗坤 樊建人[1,2] TAN Junyao;WANG Qiang;LUO Kun;FAN Jianren(State Key Laboratory of Clean Energy Utilization,Zhejiang University,Hangzhou 310027,China;Zhejiang Key Laboratory of Clean Energy and Carbon Neutrality,Hangzhou 310027,China)

机构地区：[1]浙江大学能源高效清洁利用全国重点实验室,杭州310027 [2]浙江省清洁能源与碳中和重点实验室,杭州310027

出　　处：《工程热物理学报》2024年第11期3317-3322,共6页Journal of Engineering Thermophysics

基　　金：国家自然科学基金资助项目(No.52206281);浙江省自然科学基金资助项目(No.LY24E060002);中央高校基本科研业务费专项资金资助(No.226-2024-00017)。

摘　　要：高效准确的风力机尾流建模是风电场运行优化的关键。本研究基于有限的计算流体动力学高保真数据和大量低保真数据,采用迁移学习和深度神经网络方法(TL-DNN)建立风力机高保真尾流代理模型,并利用大涡模拟结果对模型性能进行评估。结果表明,利用10组高保真数据和4000组低保真数据训练获得TL-DNN模型,其预测的流场平均相对误差为2.77%、近尾流和远尾流流向速度相对误差均小于4%,证明模型对尾流发展预测性能表现良好;模型在线计算时间仅为0.02 s,能实现风力机尾流实时预测,为风电场尾流控制提供技术支撑。Efficient and accurate modeling of wind turbine wake is crucial for optimizing the operation of wind farms.Based on limited high-fidelity computational fluid dynamics data and a large number of low-fidelity data,this study uses transfer learning and deep neural network(TL-DNN)to establish a high-fidelity wake surrogate model for wind turbines,and uses large eddy simulation results to evaluate the performance of the model.The results showed that using 10 high fidelity datasets and 4000 low fidelity datasets to train the TL-DNN model,the average relative error of predicted spanwise velocity was 2.77%,and the relative error of flow velocity was less than 4%,proving that the model has good performance in predicting wake development;Meanwhile,the online calculation time of the model is only 0.02 seconds,which can achieve real-time prediction of wind turbine wake and provide technical support for wind farm wake control.

关 键 词：风力机 尾流代理模型 迁移学习-深度神经网络 大涡模拟 

分 类 号：TK83[动力工程及工程热物理—流体机械及工程]