Model predictive regulation

www.imm.dtu.dk Diabetes Mellitus type 1 is currently a chronic disease without a cure aecting more than 180 million people worldwide. It is characterized by the inability of the pancreas to secrete insulin and results in a too high glucose concentration in the blood. The disease is treated by directly injecting the deciency subcuta-neous (sc) insulin by the patient. The aim is to maintain a glucose concentration as close to 90 mg=dl as possible. The decision of the amount of sc insulin to inject is based on a measurement of the glucose concentration and is chosen by the patient. In the fear of too low glucose concentration, hypoglycaemia, there is a tendency to injecting too little sc insulin which results in too high glucose concentration. This gives long-term complications like kidney failure, amputa-tion, blindness and nerve damage. By removing the decisions from the patient these complications can be reduced or removed. The articial pancreas is a method of removing the decision from the patient where a computer calculates the amount of sc insulin to inject. The calculations are based on measurements of glucose concentration in the blood of the patient and is taken for every e.g. 5 minutes. There has been widely research in the arti cial pancreas. In this thesis a system of a diabetes patient is modelled by the Hovorka model which is a well-known model of such. The system is regulated by a Model Predicted Controller (MPC) where a Finite Impulse Response (FIR) model is used to predict the future glucose concentration. The regulation by MPC has diculties regulating the Hovorka model in some scenarios. The FIR models lack of a method for feed-forward and the lack of a method for avoiding violating the hypoglycaemia bound makes this regulation unsucient. The feed-forward has been used in other approaches with good results where a meal detection algorithm is used to estimate the disturbance. i