Assembling neural networks for modelling of large nonlinear dynamic systems

Several feedforward neural networks with one hidden layer were trained to predict a selected coupler force in a large and highly nonlinear freight train dynamic system. Neuron activation functions were hyperbolic tangent and linear for hidden and output layers respectively. Data inputs included train control information and train control history. For more complex responses the first differences of the inputs and input histories were also used. Neural networks were successfully trained to predict coupler forces for selected positions in empty and loaded trains