A neural cluster structure with single component prediction (NCSCP) was proposed for X-ray fluorescence spectrometry in a multivariable system. The neural cluster structure is built by the collection of a group of neurons which have close relationships among one another. In X-ray fluorescence analysis, the structure is constructed by choosing the elements in which there exist serious matrix effects, and deleting the components containing large noise. The predictability of the neural cluster structure was compared with that of the classical backward error propagation algorithm with single component prediction. The results show that the neural cluster structure is significantly superior to the classical algorithm in prediction accuracy, antidisturbance and the predictabilty to outliers.