A feedforward neural network is trained on noisy data from a validated computer simulation of a 200 MW oil fired, drum-type turbogenerator unit at Ballylumford power station in Northern Ireland. Local nonlinear models, based on a multilayer perceptron with one hidden layer, are shown to give comparable predictive results to those obtained from linear multivariable ARMAX models. Neural modelling issues like the dimension of the input vector, training with noisy data, training algorithms and model validation are highlighted and discussed. A global nonlinear neural network boiler model is developed and shown to produce significantly improved predictions of the plant outputs across the complete operating range. It is concluded that neural networks can constitute a powerful tool for nonlinear modelling and identification of industrial plant.