MSc Thesis of Ioannis Sarantis:
Robust Model Predictive Control Strategies in Fuel-Cell-Car-Based Microgrids.

Abstract:
In this MSc thesis, several fuel cell vehicles, which are parked in a parking lot, are considered as the distributed generation units of a microgrid. The fuel cell vehicles can provide power to satisfy a certain load demand profile, e.g. the power demand of a hospital, and they are also able to exchange power with the main power network. In order to operate the microgrid with the minimum operational cost while at the same time satisfying the power demand, a centralized control system is considered to determine the production profile of each distributed generation unit at large time scales, e.g hours. It is assumed that any unbalanced power at small time scales, e.g. seconds, is handled by a lower-level control system embedded in each distributed generation unit.

The hybrid dynamics of the components of the fuel cell vehicles, as well as the overall system constraints, lead to the selection of model predictive control as an appropriate control method for the operation of the microgrid. The mixed logical dynamical framework for modeling hybrid systems is employed to model the fuel cell and the battery of the fuel cell vehicles, considering the different operational modes of such components. The operational limits of the distributed generation units, the limits on the power that can be traded with the main power network, and the power balance that has to be guaranteed in the microgrid, are translated into constraints. Model predictive control is able to incorporate all these constraints while it still considers the hybrid nature of the system.

The inherent uncertainty in the prediction of the power demand requires the development of robust model predictive control methods. A min-max robust model predictive control method is employed to deal with the uncertain power demand prediction. However, this method controls the system by considering the worst case of the uncertainty and therefore, it is a conservative method. The conservatism of this method requires the development of alterna- tive robust model predictive control methods. Hence, chance-constrained and scenario-based robust model predictive control methods are developed in order to reduce the conservatism of the min-max approach. The two latter methods are able to reduce the conservatism of the min-max method by taking into account more information regarding the uncertainty in the power demand, than only considering the worst case of the uncertainty as in the min- max method. The chance-constrained and the scenario-based robust model predictive control methods are able to provide a smarter scheduling of the fuel cell power production profile and can manage the stored energy in the batteries of the fuel cell vehicles more cost efficiently.

The developed robust model predictive control methods are used in a case study in which the parking lot is considered to contain fuel cell vehicles that can be used to satisfy the power demand in the microgrid and that can also exchange power with the main power network. The min-max robust model predictive control method is as expected the most expensive method for control, while the chance-constrained and the scenario-based robust model predictive control methods provide a lower operational cost for the microgrid.

MSc Thesis of Bart Kaas:
Model Predictive Control Architecture Comparison for a Power Scheduling Electric Vehicle Aggregator

Abstract:
Energy flexibility is the ability to change power production or consumption over time. It is required for a power system to function properly, to balance supply and demand. Currently, the largest providers of energy flexibility in the Netherlands can be found on the supply side and consist mainly out of fossil fueled power plants. As these are set to phase out in the near future and be replaced by mainly variable renewable energy sources, such as wind and solar, the necessity for energy flexibility on the demand side is set to be increased. Therefore, a new role is expected to arisemin the power system, namely that of the aggregator. Aggregators will combine small scale energy flexibility providers and provide this aggregated energy flexibility to the power system.

In this MSc thesis, the focus lies with a power scheduling Electric Vehicle (EV) aggregator. Considering that the number of EVs will increase in the near future, the lack of control over charging a large fleet of EVs may result in an overloaded distribution grid. The charging behavior of a large fleet of EVs, connected in a Vehicle to Grid (V2G) setting, is formalized as a Model Predictive Control (MPC) optimization problem. Allowing power consumption to be shifted within a finite prediction horizon. To optimally valorize the energy flexibility of the fleet and respect the limits of the distribution grid, the control problem is extended to include spatial information and network constraints. The goal is to develop multiple control algorithms to solve this control problem using distributed optimization.

The contributions of the work in this MSc thesis are threefold. First, the ability to optimally valorize the energy flexibility is increased by including spatial information in the distribution grid, represented as subsets of the fleet, such that congestion management services can be provided. Secondly, a parallel implementation of a coordinated distributed MPC is developed using resource allocation with feasible iterations for binary on/off input systems. Thirdly, a hierarchical MPC algorithm is developed using virtual batteries to represent the aggregated behavior of a fleet of EVs, for which new tight constraints are derived to better represent the EV fleet.

To conclude, numerical experiments are performed in closed loop to study the behavior of the developed algorithms with respect to a centralized benchmark. The experiments show that for a growing EV fleet, the hierarchical algorithm remains at the same approximate error with respect to the benchmark. This, while the distributed algorithm approaches the benchmark very well, with limited communication and in relatively short computation times with respect to the benchmark. For a growing number of subsets using the same amount of EVs, the hierarchical algorithm is able to come up with a feasible solution reasonably fast. Whereas, the distributed algorithm shows a drastic decrease in computation time as multiple smaller problems are now solved. Both algorithms achieve this at increasing costs. Future work is expected to further improve the hierarchical algorithm such that it will be able to outperform the distributed architecture in practical applications.