Long-term prediction of hydraulic system dynamics via structured recurrent neural networks

This work presents a methodology for designing neural networks to predict the behavior of nonlinear dynamical systems with the guidance of a priori knowledge on the physical systems. The traditional neural network development techniques are known to have considerable disadvantages including tedious design process, long training periods, and most notably convergence/stability problems for most real world applications. The presented approach, which circumvents such bottlenecks, is especially useful in developing efficient neural network models when full-scale models are not available. This study illustrates the application of the method on a highly nonlinear hydraulic servo-system so to estimate accurately the chamber pressures of its hydraulic piston in extended time periods.