AI-Driven HVAC Controller for Tropical Commercial Buildings
A recent study introduces a novel controller known as CQD-ERL, designed for the supervisory management of tropical water-cooled chiller plants and their air-side systems. This controller features a product archive that organizes specialized policies based on a data-driven operating context, which includes daily weather patterns and load regimes, along with a behavior descriptor that remains constant. It utilizes a gradient-free evolutionary operator alongside a soft-actor-critic policy-gradient operator that both access a shared replay buffer. Prior to execution, each action is screened by a deterministic safety shield. The training occurs in a two-tier reduced-order environment simulating the cooling-tower approach, humidity, and latent load constraints of a commercial building in Singapore. Performance is assessed through an annual backtest against the ASHRAE Guideline 36 standard. This research, aimed at enhancing HVAC control efficiency in tropical regions to lower energy usage in commercial structures, falls under Computer Science > Machine Learning and was submitted to arXiv (ID: 2608.11324).
Key facts
- CQD-ERL is a contextual quality-diversity evolutionary reinforcement-learning controller.
- It is designed for supervisory control of tropical water-cooled chiller plants and air-side systems.
- The controller maintains a product archive of specialized policies.
- Policies are indexed by operating context, weather/load cluster, and behavior descriptor.
- It uses a gradient-free evolutionary operator and a soft-actor-critic policy-gradient operator.
- A deterministic safety shield filters every action before execution.
- Training environment models a Singapore commercial building with latent load, cooling-tower approach, and humidity constraints.
- Evaluation is done via a full annual backtest against an ASHRAE Guideline 36 baseline.
Entities
Institutions
- arXiv
- ASHRAE
Locations
- Singapore