Enhanced Deep Reinforcement Learning for Balancing Energy Efficiency and Occupant Comfort

Using Reinforcement Learning (RL) to control the energy consumption while preserving the comfort level in a building has shown promising. For instance, in our previous work, we have proposed a first implementation of RL, where we have defined the energy use environment by a minimal set of parameters representing the indoor temperature, humidity and the light intensity. Although, this first implementation has shown good results, the set of states, actions and the reward function are over simplified and can be fairly improved to better represent a reel occupant behavior including energy operation and human comfort. In this paper, we extend our RL customization by defining a new state representation, new actions and new reward function that considers more descriptive indoor building features and a more elaborated comfort that considers thermal, and visual comforts and the indoor air quality. Due to the increased complexity brought by the new state space, a new variant of Deep RL is used to allow appropriate decision making when choosing action for efficient energy consumption while maintaining occupant comfort. The proposed RL customization is used to control a realistic simulated energy-use environment developed using EnergyPlus and OpenStudio. Using different scenarios of using energy with a building, the results show that our DRL was able to learn a policy that controls the simulated environment. The DRL leads to an energy saving ranging from 5% to 17% while maintaining a same comfort when compared with the thermostat Strategy.