Data-driven MIMO control of room temperature and bidirectional EV charging using deep reinforcement learning: simulation and experiments
The control of modern buildings is, on one hand, a complex multi-variable control problem due to the integration of renewable energy generation devices, storage devices, and connection of electrical vehicles (EVs), and, on the other hand, a complex multi-criteria problem due to requirements for overall energy minimization and comfort satisfaction. Both conventional rule-based (RB) and advanced model-based controllers, such as model predictive control (MPC), cannot fulfil the current building automation industry requirements of achieving system-wide optimal performance of a modern building at low commissioning and maintenance costs. In this work, we present a fully black-box, data-driven method to obtain a control policy for a multi-input-multi-output (MIMO) problem in buildings – the joint control of a room temperature and a bidirectional EV charging – with the aim to maximize occupants comfort and energy savings while leaving enough energy in the EV battery for the next trip. We modelled the room temperature and EV charging using recurrent neural networks and a piece-wise linear function, respectively, and used these models as a simulation environment for the Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm to find an optimal control policy. In the simulation, the DDPG control agent achieved on average 17 a standard RB controller. Similarly, for the joint room heating and bidirectional EV charging control, the DDPG MIMO controller achieved on average 12 savings compared to two standard RB controllers. We also validated the method on the DFAB HOUSE at Empa, Duebendorf, in Switzerland where we obtained 27 energy savings at better comfort over three weeks during the heating season.
READ FULL TEXT