RT23 | Development of a PHIL Test System with a Pre-trained AI-based Load and Charging Management

Electric mobility is a crucial measure for reducing CO2 emissions in the transportation sector, thereby mitigating the effects of climate change. While the range of choice for electric vehicles (EV) continues to rise, the possibility to recharge independently of time and location will turn out to be the EV's essential problem. The looming prospect of power grid failures and their subsequent charging restrictions is omnipresent in both the media and public discourse. To address these threats, two main measures are possible: the expansion of the energy system and grid control. The universal grid and energy supply expansion are ineffective, costly, and time-consuming, which leaves grid control as the remaining opportunity. In the research project dtec emob a future-proof intelligent grid control is developed. To prevent problems on the grid as well as on the EV side, it consists of two subsystems: a classical load management system and an EV-specialized charging management system. Both systems are laid out to suit the demands of Hamburg’s power grid. The first approach employed in the project uses a reinforcement learning algorithm with the objective of reducing the medium voltage/low voltage transformer’s congestion. The controlling agent is trained in a Python environment extended with the pandapower package for power flow calculations. To test, validate, and refine the grid control, a Power-Hardware-in-the-Loop system is built in the project’s laboratory. A Hamburg-based low-voltage reference grid containing 57 household loads (each with an EV charging station and an EV) and a distribution transformer is used. The power system simulations are performed on an OPAL-RT simulator starring ePHASORSIM, which includes a PowerFactory grid model for power-flow calculations. A single household grid connection point with its EV load is represented using an optically connected linear amplifier as a grid simulator for the power supply, a programmable load for the conventional household loads, a smart controllable AC charging station (ACC), and an EV. The agent that has been pre-trained is subsequently transferred from Python to the Simulink-based real-time environment of OPAL-RT, where it receives simulation and measurement results. It controls the charging power of each simulated EV as well as the charging power of the included power hardware using the built-in communication interfaces to limit the ACC’s power output. The current focus of the project is to transfer the pre-trained agent to its new environment and use the ACC's communication channel with the OPAL-RT's included communication outputs to control it. Further research focuses on the expansion of the PHIL system with HP-DC chargers, bidirectional charging stations, digital twins of different mobility carriers, and traffic flow simulations of Hamburg.