|
Research interests
Model Predictive Control Airborne Wind Energy Learning-based Design of Observers and of Feedback Controllers Polynomial Chaos Expansions Vehicle Stability Control Switchgear |
| Note: all the citations reported below refer to the articles listed in the publications page. |
|
Model Predictive Control (MPC)
MPC is a powerful approach to cope with system constraints while optimizing a desired performance criterion, by embedding a finite-horizon, constrained optimal control problem in the controller. In the field of MPC, we have delivered research results pertaining to the following topics: Prediction models with guaranteed uncertainty bounds ([J51]): an important aspect of MPC under uncertainty is the derivation of tight error bounds on the predicted trajectories. Along this line of research, we have recently delivered new interesting results on simulation error bounds and stability constraints in system identification, exploiting SM techniques and multi-step-ahead prediction models. We are currently investigating the use of these techniques to realize moving horizon estimators with tight estimation error bounds. Learning-based model predictive control ([J20],[J27],[J46],[J50]): we developed approaches to derive a MPC law using a model identified from data, together with a model of the related uncertainty. The initial works were aimed at nonlinear systems, where we used 1-step-ahead nonlinear Set Membership (SM) models [J20],[J27]. Recently, we delivered new results for LTI systems, exploiting a combination of 1-step-ahead and multi-step-ahead models [J46] and tube-based robust MPC design. The approach has sound theoretical properties and lower conservativeness with respect to approaches based on 1-step-ahead models only. We also developed an application in HVAC control for a large business center, where we derived a structured model from a 1.5 years data-set, and used such a model to optimally tune the feedback control law [J50]. Adaptive model predictive control ([J31], [J32]): by combining set membership identification techniques and MPC, we developed a new approach to control uncertain linear systems subject to constraints. At each time step, the set of all system models consistent with prior assumptions and measured input-output data is computed by solving a series of convex programs. Then, the input is calculated following a receding horizon strategy, by solving a robust optimal control problem that ensures constraint satisfaction for all the models in such a set. Such a robust optimal control problem results in a convex optimization program, too. We proved several theoretical properties of this approach and applied it in experiments on a quad-tank laboratory set-up, showing the advantages with respect to standard MPC and to other adaptive approaches. We also extended the approach to time-varying systems [J45]. As a side activity of this line of research, we delivered new results concerned with optimal experiment design for constrained linear systems, in a set membership framework. Stochastic model predictive control via scenario optimization ([J22], [J24]-[J25], [J33]): we developed a new technique to tackle optimal control problems in the presence of stochastic uncertainty affecting the constraints. In particular, we were among the first researchers to propose the use of scenario optimization techniques to derive approximate solutions to the chance-constrained optimal control problems arising in stochastic MPC. By proving new results in the field of scenario optimization, and by analyzing rigorously the closed-loop behavior of the system, we were able to exploit the special structure of the convex optimization problem arising in MPC to dramatically reduce the sample complexity, i.e. the number of scenarios that need to be considered in order to guarantee a desired maximum probability of constraint violation. The resulting stochastic MPC approach is able to tightly achieve the desired violation probability in closed-loop operation, with a computational complexity only slightly larger than the corresponding nominal model predictive controller. Economic model predictive control ([J26]): we studied the properties of MPC laws for nonlinear systems where the aim is to optimize a generic objective function, which does not necessarily incorporate a reference tracking term (so-called economic MPC). In particular, we proposed a new terminal constraint where, contrary to the standard approach, the terminal steady state is a free variable in the optimization problem. The resulting control law exhibits in general a feasibility set no smaller than that of an MPC law with standard terminal equality constraint, and is able to give rise to non-periodic and non-steady trajectories which might be favorable in economic problems. We derived several theoretical results concerned with the recursive feasibility and optimality of such an approach, and showed its features on an inverted pendulum example. Fast model predictive control ([J5], [J11]-[J13], [J15]): we proposed an approach to reduce the on-line computational burden of MPC, by deriving off-line an approximation of the corresponding nonlinear function mapping the state variable to the control input. By using set-membership approximation techniques, we first derived bounds on the worst-case approximation error that depend on the number of data employed to compute the approximated controller. We then used such bounds to link the number of employed data points to the properties (stability and tracking accuracy) of the closed-loop system. In subsequent contributions we studied the extension of the approach to discontinuous MPC laws and its application to vehicle stability control. In the latter, we implemented successfully the approximated controller in real-time on an embedded processor with limited computational power. ^Top of page^ |
|
Airborne Wind Energy
Airborne wind energy, or high-altitude wind energy, is the umbrella-name given to a series of technologies that are being developed by several companies and research groups around the world, whose aim is to harness wind power by using tethered wings, which can reach higher altitudes with lower construction costs with respect to conventional wind turbines. Albeit these ideas trace back to the late '70s, their industrial development has been made possible only in the last decade, thanks to advancements in many of the involved technological fields. In the field of AWE, we contributed several important results during the last eight years. We were among the first researchers to build an experimental set-up (20 kW rated power) and use it to investigate the potentials of the idea. In [J2], [J6], [J9]-[J10], [J14] and [J18] we disseminated these first results, which include the application of efficient economic MPC to stabilize the trajectory of the wing and maximize the generated energy, the optimization of the system operation, also considering multiple generating units, the estimation of the cost of the generated energy, and a study on the potentials of this technology for marine transportation. Later on, we built a new small-scale prototype specifically aimed at studying the modeling, estimation and control problems related to AWE [J36]. By using such a prototype, we delivered one of the first feedback control approaches that have been proven to achieve autonomous operation in real-world experiments for a significant amount of time [J29], together with the first experimental study on sensor fusion [J28] and the first experimental implementation of an adaptive approach to cope with uncertainty in wind direction [J35]. We also implemented the developed techniques on a prototype built in the context of the Swiss Kite Power initiative (a joint research effort between ETH, EMPA and FHNW in Switzerland) and demonstrated full power-cycle operation with such a system [J37]. In a project at ABB Corporate Research, we realized the world-first fully autonomous take-off and flight of a rigid tethered glider [J41]-[J43], developing a hierarchical and distributed control system, and designing and building the ground station we used for testing. We achieved fully autonomous take-off in just 1.5 m of space, at 4g acceleration, thus demonstrating the possibility to carry out this maneuver in very small space. In a collaboration with the group at UF Santa Catarina, we tested advanced filtering techniques for kites on extensive experimental data [J49]. ^Top of page^ |
|
Learning-based Design of Observers and of Feedback Controllers
The most common approach to design an observer and/or a controller consists of two steps: in the first one, a model of the system at hand is derived by exploiting prior assumptions and experimental data; in the second step, such a model is used to design the observer or the controller. In this research topic, we explore the potential of another conceptual approach, named direct design, in which the observer or the controller is designed in one-step, directly from prior assumption and experimental data, avoiding the explicit derivation of a model of the system. About direct observers, we proposed a technique based on set-membership identification to design the observer from experimental data, and analyzed its theoretical properties and its connections with Moving Horizon Estimation (MHE). We also proposed a mixed approach for nonlinear systems where a direct observer and a MHE are combined [J21]. We applied the approach to the problem of estimating the sideslip angle of a road vehicle, testing the algorithm both on a high-fidelity model and on real data [J30]. Moreover, we investigated the use of direct observers in a feedback control loop and the robust design of the controller, again in the context of automotive control systems [J16]. Regarding the direct design of feedback controllers, we proposed two new approaches, the first one based on the idea of deriving an inversion of the system's dynamics using experimental data (either off-line [J23] or on-line [J40]), the second one aimed at learning an existing controller (e.g. the behavior of a human operator, or an unknown control algorithm) from data collected in closed-loop operation, [J38]. The proposed techniques are based on convex optimization and yield control algorithms that are relatively simple to implement. Besides deriving, for both approaches, theoretical results concerned with closed loop stability, we successfully applied the second one to control the flight of power kites in real-world experiments. ^Top of page^ |
|
Polynomial Chaos Expansions
Simulating complex stochastic systems might be a time consuming task, especially when standard Monte Carlo approaches are used and the number of the involved random variables is relatively large. Polynomial chaos expansions represent a valid alternative to Monte Carlo simulations and are able to dramatically decrease the computation time, however, in order to use this method, the expansion's coefficients have to be derived, usually by using a reduced set of initial simulations. In this context, we proposed a new approach, based on convex optimization, to compute the expansion's coefficients by exploiting a very small number of initial simulation runs [J19]. We applied the method to test cases, including a stochastic chemical oscillator encountered in the area of systems biology, showing that also relatively large dimensions (up to around 20) of the vector of random variables can be treated effectively. ^Top of page^ |
|
Vehicle Stability Control
Active stability systems aim to improve the safety and maneuverability of passenger cars by changing their steering behavior and/or aiding the driver to keep directional stability. The most common vehicle stability control systems employ the wheel speeds, steering angle, lateral acceleration and yaw rate as feedback variables, and act by operating differentially the four brakes. Other possibilities include the use of active differentials and active or semi-active steering systems. In this topic, we delivered contributions related to the robust control design for vehicle stability systems, considering the use different actuators and of different control techniques [J1], [J3]-[J4], [J7]- [J8]. Such results have been developed in part during my work experience at Fiat Research Center near Torino, where I have been also working on fault detection and accommodation schemes for vehicle stability systems. ^Top of page^ |
|
Switchgear
In my former position at ABB Corporate Research I've been working in the area of switchgear. These devices have to accomplish an extremely challenging task: to switch from very good conductors to very good insulators in few milliseconds, in order to interrupt large prospective short circuit currents and to withstand, immediately afterwards, the system's voltage, avoiding a dielectric failure. In addition, they have to work reliably and efficiently even after tens of years, eventually in harsh environments, and they have to be cost-competitive. The phenomena associated with the switching are extremely complex, not fully understood, and they involve multi-physical, multi-scale problems, requiring competences in plasma physics, mechanics, and materials. ABB is a global leader in switchgear systems and has delivered many breakthroughs, building a world-class know-how of such complex phenomena. My research activities in this field involve modeling and identification of linear-parameter-varying dynamical systems encountered in plasma physics [J39]. The full details of such research activities cannot be disclosed due to confidentiality clauses. ^Top of page^ |