Self-Learning Hybrid Neural Model Predictive Control for Uncertain Nonlinear Systems with Real-Time Adaptive Optimization

Uncertain nonlinear systems are common in industrial automation, chemical processing, energy conversion, and advanced manufacturing, where nonlinear dynamics, parameter variations, modeling errors, and external disturbances can significantly degrade control performance. This study proposes a self-learning hybrid neural model predictive control (SH-NMPC) framework integrating online neural system identification, constrained predictive optimization, and real-time adaptation. The system identifier employs a multilayer feedforward neural network with two hidden layers of 10 neurons each, hyperbolic tangent activation functions, and a linear output layer. Network weights and biases are updated online using a gradient-based adaptation rule driven by the instantaneous prediction error, enabling continuous refinement of the internal model under changing operating conditions. The updated neural model is embedded within the MPC prediction mechanism to forecast system behavior and determine optimal control actions by minimizing a constrained cost function incorporating tracking error, control effort, and input variation. Closed-loop stability under online adaptation is investigated using a Lyapunov-based analysis, establishing sufficient conditions for bounded prediction errors, parameter-estimation errors, and control signals under bounded uncertainties and disturbances. The framework is implemented in MATLAB/Simulink and evaluated on a benchmark uncertain nonlinear system subjected to parameter variations and external disturbances. Performance is benchmarked against conventional PID, standard MPC, adaptive MPC, and neural-network-based MPC using tracking accuracy, settling time, overshoot, disturbance rejection, control effort, and robustness criteria. Simulation results demonstrate improved tracking, faster adaptation, enhanced disturbance rejection, and robust closed-loop behavior. Computational feasibility is assessed by quantifying neural-update and MPC-optimization execution times relative to the controller sampling interval, confirming that the complete control cycle satisfies the prescribed real-time constraint. The proposed SH-NMPC therefore provides a stable, adaptive, robust, and computationally feasible solution for intelligent control of uncertain nonlinear systems.