Roll force and torque prediction play a critical role in modern rolling schedule control and optimisation. For a given steel grade, the roll force and torque will be determined by the stock temperature, mill reduction, roll speed, friction and heat transfer, etc. A finite element (FE) model is first developed for the modelling of the rolling process, and gives prediction of stock temperature, strain, strain rate, and stress profiles during rolling. The average roll force and torque were then calculated by post-processing based on the distribution of the local variables of strain, strain rate, temperature, etc. A series of rolling experiments has been conducted to validate the developed FE model before the FE model is then used to generate the training data for developing the neural network (NN) models. An important innovation here is the integration of orthogonal design techniques for training data generation at minimal computational cost with the help of a validated FE model. This effectively overcomes the difficulty of obtaining good training data with sufficient excitations and balanced distribution via data collection. Novel techniques have been adopted in the NN modelling, including structure optimisation and double-loop iterative training to achieve an accurate and robust NN model based on the existing process data. The resulting NN models are feasible for on-line control and optimisation of the rolling process, which is impossible for the FE model due to the heavy computation involved. Also, ensemble modelling techniques have been adopted to improve the model prediction, which can also give confidence bounds for the predictions. The developed NN models can be easily extended to cover different steel grades and stock sizes in a straight-forward way by generating the corresponding training data from an FE model.