基于深度神经网络构建风电机组性能模型的超参数选择

SELECTION OF HYPERPARAMETERS FOR WIND TURBINE PERFORMANCE MODEL BASED ON DEEP NEURAL NETWORK

风电行业中基于深度神经网络技术来构建风电机组性能模型受到了业内人员的广泛关注,然而对于深度神经网络自身超参数选取对模型结果的影响仍缺少系统性分析。以深度神经网络隐含层层数、深度神经网络形状及隐含层神经元数量为研究基础,研究了超参数取值对风电机组性能模型构建效果的影响,并通过实际运行数据验证了参考风电机组的最优配置及其推广性。结果表明,深度神经网络隐含层层数、深度神经网络形状及隐含层神经元数量这3个超参数的不同配置对最终构建的风电机组性能模型存在一定影响,但从工业应用的角度来看,这些影响基本可以忽略;同时,基于参考风电机组得到的最优超参数配置能够直接推广到同一风电场相同型号的所有风电机组上应用。

The construction of wind turbine performance models based on deep neural network technology has attracted widely attention in the wind power industry.However,there is still a lack of systematic and indepth research on the influence of hyperparameter selection in neural networks on the performance model.This paper attempted to study the effect of the hyperparameters values on the wind turbine performance model from three aspects of the number of hidden layers,the shape,and the number of hidden neurons of the deep neural network,to verify the optimal configuration and generalizability of the reference wind turbine through actual operating data.The results show that the three deep neural network hyperparameters have a certain impact on the final turbine performance model under different configurations.However,from the perspective of industrial applications,they can be basically ignored,and the optimal hyperparameter configuration based on the reference wind turbine is suitable to be promoted to all the same type wind turbines in the same wind farm.