Dynamical system simulation with attention and recurrent neural networks

Accurate and efficient real-time simulation of nonlinear dynamic systems remains an important challenge in fields such as robotics, control systems and industrial processes, requiring innovative solutions for predictive modeling. In this work, we introduce a novel recurrent neural networks (RNN) architecture designed to simulate complex nonlinear dynamical systems. Our approach aims to predict system behavior at any time step and over any prediction horizon, using only the system’s initial state and external inputs. The proposed architecture combines RNN with multilayer perceptron and incorporates an attention mechanism to process both previous state estimates and external inputs. By training without teacher forcing, our model can iteratively estimate the system’s state over long prediction horizons. Experimental results on three public benchmarks show that our method outperforms other state-of-the-art solutions. We highlight the potential of our proposal for modeling and simulating nonlinear systems in real-world applications.