Instruments and Systems: Monitoring, Control, and Diagnostics Annotation << Back Learning Nonlinear Predictive Control Models from Data: a Comparison of Artificial Neural Networks and NMPC for Real-Time Control of Nonlinear Dynamic Systems V. KAFA, D.A. GAVRILOV, E.A. TATARINOVA, I. MURHIZH, N.N. SHCHELKUNOV, E.O. SAVTSOV This study investigates the feasibility of replacing Nonlinear Model Predictive Control (NMPC) with a feedforward artificial neural network (ANN) for real-time control of complex, nonlinear, time-variant systems. While NMPC is effective, it requires precise models and significant computational resources, challenging its application in robotics and UAVs. This research leverages historical data to train an ANN to replicate NMPC performance with reduced computational overhead. Simulations of a 2-axis gimbal system generate comprehensive datasets for training and fine-tuning the neural network. An incremental tuning strategy optimizes hyperparameters, balancing computational efficiency and control accuracy. Performance assessments across various network configurations and dataset sizes identify an optimal setup. Results demonstrate that a well-trained ANN on a comprehensive dataset can achieve real-time control with significantly lower computational demands, mimicking NMPC's effectiveness while adhering to real-time constraints. The study highlights the potential of neural networks as a viable alternative to NMPC, offering substantial improvements in computational efficiency for resource-constrained platforms in dynamic environments. Keywords: Nonlinear Model Predictive Control, artificial neural network, Real-time control, Data-driven control, Nonlinear System dynamics, 2-axis gimbal. DOI: 10.25791/pribor.8.2025.1602 Pp. 01-11.