A novel iterative learning control formulation of generalized predictive control

A novel iterative learning control formulation of the long-range predictive control method generalized predictive control (GPC) termed GPC with learning (GPCL) is developed. GPCL is applicable to processes that are repetitive in nature and are subject to a partially repeatable disturbance. It improves the regulation performance of GPC by learning and compensating in advance for the repeatable portion of the disturbance over a series of trials. GPCL is proven to have the same stability properties as GPC, and to converge asymptotically on condition that its feedback loop is stable. According to a simulation study, for six plants of varying control difficulty, GPCL outperformed GPC after two trials, and reduced the final value of the output variance by 48% on average for a 50% repeatable disturbance. In robotic edge-following experiments with a 47% repeatable disturbance, GPCL reduced the average output variance by 32% relative to GPC after three trials. After ten trials, the reduction was 43%.