Dynamic modelling of the activated sludge process: Improving prediction using neural networks