Artificial Neural Network(ANN) is a commonly used method to build process models

and is specially useful in the processes with complicated mechanism

which is difficult to be modeled by the traditional methods. The ANN method needs a lot of data. However

practically measured data generally have larger error which may lead to imprecise models built by ANN. Usually the data have to be selected before they are used to train a network

but that is not easy. In this paper a robust evaluation function method for Back Propagation Neural Network(BPNN) is proposed. In this way commonly used least square evaluation function form is changed into some special robust evaluation function form which can make the network automatically choose better data. The larger error of data is

the less use of the data is. Therefore the overfilling problem of BPNN can be belter solved so that the precision of the built model will be higher. Based on the sufficienl condition of constructing a robust evaluation function presented in this paper

differenl robusl evalualion functions can be produced. A mathematical example and a practical case study of pressure drop model of hydrocracking trickle bed reactor show that robusl evalualion functions have an obvious superiority in raising model precision and decreasing overfitting of BPNN over the least square evaluation function.