Identification of nonlinear dynamic systems using modified DRNNs

In order to apply the better diagonal recurrent neural network (DRAW) to nonlinear dynamic system identification, three different modified DRNNs are proposed and compared DRAW has more dynamic mapping capability than feedforward neural network (FNN). Meanwhile, it has simpler structure and needs less training time than full recurrent neural network (FRNN). To overcome the insufficiency of the time variable character of the weight vectors in previous research, this work modifies the training algorithm and applies to the three DRAW models including DRAW, higher order DRNN (HDRNN) and quasi DRNN (QDRNN). Meanwhile, both MIMO and SISO nonlinear dynamic systems are tested. The simulations show that the three models using modified algorithm have better performance than original models. Modified QDRNN has the best performance.