Data-driven adaptive model predictive control using dualpath-itransformer for nonlinear industrial processes

Abstract

To improve control accuracy and robustness for complex industrial processes with nonlinearities, large time delays, and time-varying operating conditions, this paper proposes a closed-loop adaptive model predictive control (AMPC) framework based on DualPath-iTransformer. In the prediction stage, a dual-path multivariate time-series model is constructed to capture both inter-variable coupling and temporal dynamic evolution. In the control stage, the multi-step prediction sequence is converted into compact predictive features, and actuator increments are directly computed via an analytical robust control law that integrates disturbance feedforward, deviation compensation, trend correction, and constraint mapping. Meanwhile, the Sparrow Search Algorithm is employed for slow-time-scale parameter optimization to reduce online computational burden. In the updating stage, a performance monitoring mechanism is introduced to trigger model fine-tuning and controller parameter re-optimization when closed-loop performance degrades, after which the updated modules are redeployed online. In this way, the proposed framework aims to alleviate the model-mismatch problem commonly encountered in conventional MPC. Experiments on representative industrial control scenarios, including main steam temperature regulation in thermal power plants, show that the proposed method achieves competitive overall performance compared with PID, conventional MPC, and LSTM-MPC baselines, while exhibiting good tracking accuracy, disturbance accommodation, and engineering deployment potential. These results suggest that the proposed framework provides a feasible data-driven approach for complex industrial process control.