为了研究人工神经网络在压气机性能曲线拟合中的应用，分别利用BP神经网络、RBF神经网络、极限学习机以及BP-GA神经网络对某微型燃气轮机压气机的性能映射关系进行模拟，分析了不同网络模型在压气机特性曲线拟合上的优劣，以及样本容量对不同神经网络模型性能的影响。结果表明：BP-GA神经网络模型不仅收敛速度快，而且精度高；相比传统BP神经网络模型，其平均绝对百分比误差可控制在0.189%以内，训练时间可缩短至19.07 s；当样本容量较少时，传统BP神经网络模型不再适用，而基于遗传算法的BP-GA模型仍然保持较高的精度。

In order to study the artificial neural network application in the compressor performance curve fitting, the mapping relationship of a certain micro gas turbine compressor performance was simulated by using BP neural network, RBF neural network, extreme learning machine and BPGA neural network respectively,the advantages and disadvantages of different network models in compressor characteristic curve fitting and the influence of sample size on the performance of different neural network models were analyzed. The results show that the BPGA neural network model not only has fast convergence speed, but also has high precision. Compared with the traditional BP neural network model, the average absolute percentage error can be controlled to no more than 0.189%, and the training time can be shortened to 19.07 seconds.When the sample size is small, the traditional BP neural network model is no longer applicable, while the BPGA model based on genetic algorithm still maintains a high accuracy.

TK472

国家科技重大专项（2017-I-0002-0002）