Modeling and Control of Nonlinear Systems Using an Improved Adaptive Moment Optimization Approach

ABSTRACT A modified diagonal recurrent neural network (MDRNN) model is presented in this paper as a model that allows the modeling and control of nonlinear dynamic systems. In practical applications, neural network models may experience challenges such as slow convergence and limited generalization performance when applied to complex nonlinear dynamic systems. These limitations have been overcome via improvements to the original diagonal recurrent neural network (DRNN), in addition to enhancing adaptation capability and learning performance. Moreover, this paper also presents another better optimization strategy, which is the improved adaptive moment optimization (IAdam) strategy, an enhanced structure of the traditional Adam (adaptive moment estimation) optimization strategy. The proposed IAdam optimization method is able to dynamically adjust the weight variables of the proposed model so that it can be trained efficiently, and the model can be more adaptable under various circumstances in the system. The MDRNN model is analyzed in detail in many nonlinear dynamical systems, where it outperforms the standard neural network models both in terms of its accuracy and convergence rates. An indirect control scheme is used to implement an MDRNN‐based controller to operate a nonlinear dynamical system. The dynamics of the proposed controller are simulated by varying the parameters and using a disturbance signal to determine the ability of the controller to recover. The simulation results show that our methodology offers both an accurate and consistent method of identification and control of complex dynamical systems, and therefore, it is a very useful tool in real applications of control systems, signal processing, and time series forecasting.