Data-Driven Condition-Aware Adaptive Model Predictive Control for Industrial Ethylene Cracking Furnaces

During long-term operation of ethylene cracking furnaces, coke deposition and operating condition variations continuously alter process dynamics, resulting in strong non-stationarity and reduced control performance of conventional model predictive control (MPC). To address this issue, a condition-aware adaptive MPC (CA-AMPC) framework is proposed, in which operating conditions are incorporated into the control loop to enable adaptive prediction and control adjustment. A segment-wise operating condition identification method combining change point detection and clustering is developed to capture both abrupt transitions and gradual process evolution. Based on the identified conditions, condition-dependent prediction models and optimized control parameters are dynamically activated to improve control adaptability under varying operating regimes. Industrial case studies demonstrate that the proposed framework achieves yield improvements of approximately 1.0%, 1.5%, and 0.3% under normal-load early-coking, normal-load severe-coking, and post-decoking high-load conditions, respectively. The total online computation time remains well below the 120 s DCS sampling interval. The proposed framework provides an effective solution for adaptive control of non-stationary industrial processes.