Neural network‐based offset‐free model predictive control for nonlinear systems

This paper proposes an offset‐free model predictive control (MPC) framework for nonlinear systems modeled using neural network‐based nonlinear autoregressive models with exogenous inputs (NARX). To address plant‐model mismatch and ensure offset‐free tracking, the NARX model is augmented with an integrating disturbance model, resulting in an extended state‐space suitable for MPC. A nonlinear observer is developed to estimate both system and disturbance states in real time. The impact of training data quality on control performance is examined through two modeling scenarios: one with rich excitation data and another with limited excitation data, reflecting practical constraints. For both cases, offset‐free MPC controllers are designed using the proposed framework. The approach is validated through simulations on a nonlinear chemical reactor and compared with a benchmark NARX‐based offset‐free MPC method employing bias correction from output prediction errors. Results show that the proposed method improves tracking performance, particularly when training data are limited.