Compressor characteristic maps from Compressor performance modelling method based on support vector machine nonlinear regression algorithm

To overcome the difficulty of having only part of compressor characteristic maps including on-design operating point, and accurately calculate compressor thermodynamic performance under variable working conditions, this paper proposes a novel compressor performance modelling method based on support vector machine nonlinear regression algorithm. It is compared with the other three neural network algorithms (i.e. BP, RBF and Elman neural networks) from the perspective of interpolation and extrapolation accuracy as well as calculation time, to prove the validity of the proposed method. Application analyses indicate that the proposed method has better interpolation and extrapolation performance than the other three neural networks. In terms of flow characteristic map representation, the root mean square error (RMSE) of the extrapolation performance at higher and lower speed operating area by the proposed method is 0.89% and 2.57%, respectively. And the total RMSE by the proposed method is 2.72%, which is more accurate by 47% than the Elman algorithm. For efficiency characteristic map representation, the RMSE of the extrapolation performance at higher and lower speed operating area by the proposed method is 2.85% and 1.22%, respectively. And the total RMSE by the proposed method is 1.81%, which is more accurate by 35% than the BP algorithm. Moreover, the proposed method has better real-time performance compared with the other three neural network algorithms.针对仅包含设计工况点的不完整压缩机特性图所带来的难题，以及精准计算变工况下压缩机热力学性能的需求，本文提出一种基于支持向量机（support vector machine）非线性回归算法的新型压缩机性能建模方法。本文从插值与外推精度、计算时长两个维度，将该方法与另外三种神经网络算法（即反向传播（Back Propagation, BP）神经网络、径向基函数（Radial Basis Function, RBF）神经网络及Elman神经网络）进行对比，以验证所提方法的有效性。应用分析结果表明，相较于其余三种神经网络算法，本文所提方法具备更优异的插值与外推性能。在流量特性图表征方面，本文所提方法在高、低转速工况区域的外推性能均方根误差（root mean square error, RMSE）分别为0.89%与2.57%；其总均方根误差为2.72%，相较Elman算法精度提升47%。在效率特性图表征方面，本文所提方法在高、低转速工况区域的外推性能均方根误差分别为2.85%与1.22%；其总均方根误差为1.81%，相较BP算法精度提升35%。此外，相较于其余三种神经网络算法，本文所提方法还具备更优异的实时性。