DeepMind’s Latest AI Research Reduces Energy Usage For Cooling Buildings

HVAC (heating, ventilation, and air conditioning) accounts for a sizable portion of the world’s CO2 emissions. Around 10% of the entire electricity demand for the world is accounted for by space cooling alone. Therefore, improving HVAC system efficiency can be crucial for mitigating climate change. As HVAC data collecting and management technologies become more prevalent, data-driven, autonomous, real-time decisions at scale are becoming an increasingly alluring way to boost productivity.

New research by DeepMind employed reinforcement learning (RL), drawing on previous work regulating the cooling systems of Google’s data centers, to increase the energy efficiency of HVAC control in two commercial buildings.

The researchers think that reinforcement learning is a good solution for HVAC control issues for a number of reasons:

1. When it comes to HVAC control, decisions must be made about when to switch equipment on and off and how hard to run each piece of equipment. 
2. To keep the occupants happy and guarantee safe system operations, other limitations must be satisfied in addition to a natural reward function. Widely used building management systems (BMS) can provide the data needed to train an RL agent, and increasingly common cloud-connected BMS can be used to provide automated supervisory control. Since decisions might have long-term consequences, there is also a crucial sequential decision-making component. 
3. Unlike Model Predictive Control (MPC), RL does not demand the creation, validation, and maintenance of an extensive and thorough physics-based model for each and every building.

The two main data sources were historical data gathered by the SOO and current data gathered by BCOOLER while it was in charge of the facility. Less than a year’s worth of facility data from the SOO in charge of the system make up the historical data. On the other hand, the AI control data is rich in exploration information that covers a variety of activities and states.

The team faced various difficulties, from standard ones like pricey and noisy data to more unusual ones like having many operational modes and multi-timescale dynamics. They used a combination of general RL solutions and domain-specific heuristics to address these problems.

The resulting system demonstrated a 9–13% reduction in energy utilization while satisfying system restrictions compared to heuristics–based controllers offered by Trane.

To be more confident in assessing the agent’s performance before deployment, the team leveraged domain knowledge to create unit tests that the action value function should adhere. They disguised distinct actions based on the environment’s state, enabling a single agent to manage several weather-dependent modes with various action spaces and constraints.

Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.