Intelligent management of mobile robot fleets for electric vehicle charging

The increasing spread of electric vehicles requires a flexible and efficient charging infrastructure. Mobile charging robots can be a useful addition to stationary charging stations, as they increase flexibility and improve user convenience. However, controlling and coordinating these charging robots is a complex optimization task, as both the charging power and the assignment of the robots to the vehicles must be taken into account. Nevertheless, there are hardly any strategies for the operation of mobile charging robots in the literature to date. This thesis investigates the use of model predictive control (MPC) and reinforcement learning (RL) to optimize charging management with mobile charging robots in this context. For this purpose, a simulation environment was developed in the Python programming language, which enables a realistic representation of a parking lot with the help of the Gymnasium Toolbox. The modeling includes a network of parking spaces, electric vehicles, stationary charging stations and mobile charging robots. These robots inherit characteristics of electric vehicles and charging stations, respectively. The local electricity grid with variable electricity prices, building loads and a photovoltaic system are also taken into account. To implement the control strategies, the RL agent was trained using the RLlib software framework, while the MPC optimization was modeled using the Pyomo toolbox and solved using the Gurobi solver. The challenge of the RL approach lies in the decomposition of the problem in order to adequately deal with the large observation and action space. In contrast, the MPC approach focuses on the efficient formulation of the optimization problem and the integration of a suitable prediction. A rule-based controller is used as a reference for the two approaches. The evaluation is carried out over four weeks spread over the year as an example in order to reflect seasonal differences in electricity generation and demand. The rule-based reference approach achieves the highest gains but the lowest user satisfaction across all scenarios. The profits of the other two approaches are slightly lower, mainly due to the lower amount of charged energy. In two of the four weeks, the MPC approach achieves a higher economic profit than the RL approach, while the RL approach delivers a better result in the winter simulation study. In the fall scenario, the economic gains of the RL and MPC approaches are similar. Overall, the RL approach shows certain advantages compared to the MPC approach in terms of user satisfaction. In summary, this thesis examines charging management strategies for electric vehicles, taking into account mobile charging robots. The focus is on maximizing profit and user satisfaction.