Exercise can induce rapid, delayed, and nocturnal glucose changes that complicate glucose forecasting and hypoglycemia detection in people with type 1 diabetes. We evaluated a multimodal dual-head multitask framework in which a shared encoder forecasts the future continuous glucose monitoring trajectory and separately classifies whether glucose falls below 70 mg/dL within 30-, 60-, or 120-minute prediction horizons. Inputs included glucose, insulin, carbohydrate intake, heart-rate, and physical-activity signals from 404 participants in the T1DEXI dataset. DLinear, iTransformer, PatchTST, and TimeXer backbones were compared using participant-disjoint training, validation, and test splits, with participant-level forecasting evaluation and statistical comparisons. Forecasting error increased with prediction horizon. The lowest overall RMSE was 13.49 mg/dL at 30 min and 21.96 mg/dL at 60 min for iTransformer, and 32.62 mg/dL at 120 min for TimeXer. Hypoglycemia detection showed a consistent sensitivity–alarm-burden trade-off. iTransformer achieved overall recall of 0.87, 0.84, and 0.81 across the three horizons, whereas DLinear produced the lowest delivered false-alarm burden of 5.12, 10.36, and 29.37 alarms per participant-day. Predict-then-threshold maintained near-perfect specificity and fewer than one delivered false alarm per participant-day, but had substantially lower recall. Single-task detectors increased sensitivity but could produce unfavorable alarm burdens, while dual-head multitask learning often provided a more moderate operating profile without being uniformly superior. These retrospective results demonstrate that glucose forecasting and hypoglycemia detection provide complementary information.
This paper investigates a tube-based robust economic linear parameter-varying (LPV) model predictive control for the optimal operation of pressurized water distribution networks (WDNs). A generalized network formulation incorporating nonlinear hydraulic head loss is developed to enhance the physical fidelity of the control model. In order to mitigate the resulting computational complexity, a LPV approximation is introduced, achieving an effective trade-off between accuracy and computational efficiency. Demand uncertainties are further addressed through a robust control design that ensures reliable system performance under varying consumption patterns. Moreover, the proposed approach determines which actuators should be activated and how their operations should be coordinated to achieve minimal operational cost, by optimally scheduling the pumps and valves to minimize costs while satisfying hydraulic and operational constraints. Finally, the effectiveness of the proposed strategy is demonstrated through simulation results conducted on the Richmond WDN.
This paper proposes a robust dissipative boundary control for a system described by a 2-D Navier–Stokes equation with polytopic uncertainties under the port-Hamiltonian formulation. The goal of the controller is to obtain a fluid behavior that suppresses the vorticity of the waves. A complete robustness analysis, including both robust stability and robust performance, is presented. Additionally, a topological analysis of the polytopic uncertainties is performed to elucidate the open-loop system behavior. Meanwhile, the dissipative boundary-based controller is obtained by selecting an appropriate storage function. As a case study application, a wave tank is used to illustrate the approach proposed in the paper.
This paper develops a design criterion for continuous-time extended state observers (ESOs) that quantifies the estimation-error energy induced by step-like disturbances under stochastic uncertainty. For absolutely continuous disturbance variations, the induced estimation-error energy is shown to admit an upper bound characterized by the H2 norm of the transfer from the disturbance variation to the estimation error. A step disturbance is then treated as the limit of a sequence generated by an approximation identity, and the corresponding finite-horizon estimation-error energy is characterized exactly in this limit. The corresponding long-horizon limit is shown to be characterized by the H2 norm of the transfer function from the disturbance variation to the augmented-state estimation error. The analysis is further extended to an Ito-type stochastic system with process and measurement noise, where the finite-horizon expected estimation-error energy is decomposed into a disturbance-induced mean-response term and a noise-induced covariance term. Based on this finite-horizon characterization, a long-horizon approximate design criterion consisting of an H2 term and a steady-state covariance term is introduced. The observer gain minimizing the approximate criterion is shown to be characterized by an algebraic Riccati equation. Numerical simulations validate the proposed limiting finite-horizon characterization by comparing analytical values with Monte Carlo sample means and illustrate the approximate criterion as a design-oriented surrogate.
