Data-Driven Robust Control for Type 1 Diabetes 2251) Meals As Expected 2) Outliers 3) Late Meals
U. sets RandomDG(mmol/min)(a) (b) (c)
Target BG/Basal IIR HCL Robust PerfectBG (mmol/L)(^0050100) Time (min) 150 200250300
5
10
15
(d)
0 50 100150200 250 300
Time (min)
0
5
10
15
(e)
(^0050100) Time (min) 150 200250300
5
10
15
(f)
ι
(mU/min) 0 50 100 150200 250
Time (min)
0
50
100
150
200
(g)
0 50 100 150200 250
Time (min)
0
50
100
150
200
(h)
0 50 100 150200 250
Time (min)
0
50
100
150
200
(i)
Fig. 3.One-meal, 300-minute experiments (50 repetitions).Top:uncertainty sets and
random realizations of parameterDG(rate of CHO ingestion).Middle:BG profiles
(with solid black lines indicating the normal BG range).Bottom:synthetized insulin
therapies. Thick solid lines indicate average BG/insulin values, and are surrounded by
an area spanning±1 S.D. In the table, we highlight in bold the best value of each
index between the robust and the HCL controllers.
2.Light activity, where parameters stay as in the previous phase except for
02 = unif(15,45).
Results, reported in Fig. 4 , evidence that both the robust and the HCL controller
can maintain BG within very tight ranges, as confirmed by theBGminand
BGmaxindicators. BG profiles are almost indistinguishable from the ideal ones
(i.e. those of the perfect controller) and for 100% of the times within healthy
ranges. Note that both controllers correctly reduce the insulin therapy below
the basal level to counteract the decrease of BG due to exercise. Hence, the
negative values of
∑
ι. The main difference is that the robust controller, due
to the superior predictive capabilities, is more timely in cutting insulin therapy
than the HCL controller, leading to a smaller excursion from the target BG
value.
Resalat et al. [ 28 ] realized a similar scenario to test their dual-hormone MPC
(300-minute simulation with a 45-minute exercise at fixed 02 =60andMM=
0 .8). While we use their same plant model, their MPC design is different in two