Model Based Reinforcement Learning Approach to HVAC optimal control

Repo structure

agent

• Multi-Objective MDP formulation with objectives as thermal comfort and energy consumption
• Lagrangian dual reinforcement learning approach
• fine tuning left to do

mask-agent

• Single objective MDP of energy consumption, and thermal comfort enforced through hard constraint
• action bound approach
• in progress: Inferring change of environment to adjust the mask accordingly

base

• single objective reinforcement learning formulation (electric cost), with demand response
• demand response, Toronto Hydro electricity ToU (time of use)
• To reduce HVAC actuation load, CAPS action smoothing utilized

Algos implemented

• Rainbow DQN
• DQN
• SAC
• PPO