The dryness of turbine's exhaust steam can hardly be obtained via direct measurements in online performance calculation of thermal system,for the steam is usually in wet state.The artificial's neural net-works was established to calculate the exhaust enthalpy of steam turbine.By analyzing the main factors af-fecting the exhaust enthalpy and transforming the original data into non-dimensional parameters,the calcu-lation speed and accuracy were compared by using different training functions in Back-Propagation (BP) neural network.Meanwhile,the accuracy of the BP Neural Network and the Radial Basis Function (RBF) network was compared as well.Consequently,when calculating the exhaust enthalpy,the BP neural net-work depends strongly on training functions,some of which will result in strong random and low calcula-tion speed.After the research,three of the training functions:'traingdx','trainscg'and'trainoss',have the ad-vantages of both fast calculation speed and high accuracy.The RBF network has the advantages of faster calculation speed and weaker dependence on training functions but the disadvantage of lower accuracy,com-pared with the BP neural network.The error will be diminished when enough training samples are provid-ed.%在线机组热力系统性能计算中,汽轮机的排汽通常处于湿蒸汽区,排汽干度目前无法实现直接测量。对此,将神经网络方法应用于汽轮机排汽焓的估算,通过分析汽轮机排汽焓的影响因素,并对数据进行无量纲化处理,对BP神经网络在不同训练函数下的计算精度与速度,以及BP神经网络与RBF神经网络计算排汽焓的准确度进行比较。结果表明:BP 神经网络对训练函数的依赖程度较大,部分函数在计算中随机性较强、计算时间较长;traingdx、trainscg和trainoss 3个函数计算时间较短、计算精度较高,可作为训练函数;RBF神经网络的计算误差较BP神经网络大,但其自适应能力强,对训练函数的依赖程度较小,在训练样本足够多时,可以减小其计算误差。
展开▼