An on-line prediction scheme combining the Karhunen-Loeve expansion and a recurrent neural network for a wall-cooled fixed-bed reactor is presented. Benzene oxidation in a pilot-scale, single tube fixed-bed reactor is chosen as a working system and a pseudo-homogeneous two-dimensional model is used to generate simulation data to investigate the prediction scheme presented under randomly changing operating conditions. The scheme consisting of the K-L expansion and neural network performs satisfactorily for on-line prediction of reaction yield and bed temperatures.