Simulation Studies Comparing Feedback Predictive Control to Model Predictive Control for Unmeasured Disturbances in the Artificial Pancreas Application

Abstract

The artificial pancreas (AP) is an electro-mechanical device to control glucose (G) levels in the blood for people with diabetes using mathematical modeling and control system technology. There are many variables not measured and modeled by these devices that affect G levels. This work evaluates the effectiveness of two control systems for the case where critical inputs are unmeasured. This work compares and evaluates two predictive feedback control (FBC) algorithms in two unmeasured input studies. In the first study, the process is a dynamic transfer function model with one measured input variable and one unmeasured input variable. The process for the second study is a diabetes simulator with insulin feed rate (IFR) measured and carbohydrate consumption (CC) unmeasured. The feedback predictive control (FBPC) approach achieved much better control performance than model predictive control (MPC) in both studies. In the first study, MPC was shown to get worse as the process lag increases but FBPC was unaffected by process lag. In the diabetes simulation study, for five surrogate type 1 diabetes subjects, the standard deviation of G about its mean (standard deviation) (i.e., the set point) was 133% larger for MPC relative to FBPC. For FBPC, its standard deviation was less than 10% larger for unmeasured CC versus measured CC. Thus, FBPC appears to be a more effective AP control algorithm than MPC for unmeasured disturbances and may not perform much worse in practice when CC is measured versus when it is unmeasured since CC can be very inaccurate in real situations.