Data-Driven Warm-Start for Real-Time Deep Learning Model Predictive Control in Rotary Kilns

In industrial process control systems, achieving accurate and adaptive model predictive control (MPC) remains a significant challenge due to the inherent nonlinearity and time-varying dynamics. Deep learning-based MPC (DL-MPC) offers strong modeling capabilities but often suffers from high computational cost and poor real-time performance, limiting its deployment in latency-sensitive applications. To address these challenges, this paper introduces a data-driven warm-start MPC framework (DDWS-MPC) designed to enhance computational efficiency in industrial settings. The proposed framework integrates key components: (1) an offline-trained MPC approximation model that generates high-quality initial control solutions from system states, (2) a trust-region constrained convex optimization module performing Jacobian-guided online residual refinement from the Uinit warm-start. Experimental validation on a zinc oxide rotary kiln demonstrates 38 % efficiency improvement under 100-iteration limits and 71 % iteration reduction at equivalent control precision. These results demonstrate the effectiveness of DDWS-MPC in balancing computational efficiency with control performance, thereby providing a practical and deployable solution for real-time DL-MPC in latency-sensitive industrial environments.