Cooling buildings is expensive both in terms of electricity bills and carbon emissions. In India alone, buildings consumed 1,532 TWh of electricity in 2024. Nearly half of that went to heating and cooling. The HVAC systems in large hospitals, offices, and residential towers are major contributors.
But here's the thing: HVAC systems don’t usually run efficiently. They’re complex, dynamic, and often managed with a mix of static rules, manual overrides, and outdated assumptions. That’s a lot of wasted energy.
So, our team behind DeJoule decided to take a different approach. DeJoule is a Full-stack Building Management System (BMS) designed to eliminate energy waste and boost building performance, every operational minute. It combines control, data intelligence, and automation into one platform. Central to that is a Reinforcement Learning (RL) engine that optimizes HVAC systems in real time.
This isn’t a lab experiment. DeJoule’s RL system has been deployed across 30+ hospitals in India. Here’s how it works, and what it’s achieving.
Why traditional control falls short
Before Reinforcement Learning, the team tried two common approaches:
1. PID controllers
These are feedback systems that adjust one variable at a time. They're simple but don’t scale well to complex systems where components constantly affect each other, and they require manual tuning.
2. Custom scripting with If-Then rules
Using a purpose-built scripting language called Joule Recipes, facilities teams could create basic automation rules. But the logic was either too simplistic or too hard to manage at scale. Each site required custom development, and any changes meant more effort from engineers.
Both methods hit a wall. Our team needed something adaptive, scalable, and smart enough to handle real-world complexity. That’s where DeJoule’s Reinforcement Learning came in.
What makes Reinforcement Learning a better fit
Reinforcement Learning doesn’t rely on historical data or labeled examples. Instead, it learns by doing. It tests different actions, measures the outcome, and gradually figures out the best decisions to maximize long-term results.
In this case, the RL engine inside DeJoule controls the chiller plant, the part of the HVAC system that does the heavy lifting in cooling. It learns how to minimize energy use while keeping building occupants comfortable. No static rules. No constant tuning.
Why it works so well:
- No historical data needed: The RL system can start fresh in new buildings where there's no prior operational data.
- Edge-ready: It runs on small, local devices. That means real-time decisions without relying on cloud connectivity.
- Self-tuning: The system adapts to changing loads, weather, and equipment condition automatically.
- System-level intelligence: It doesn’t just optimize one piece of equipment, it balances all interdependent components.
Inside the DeJoule RL Engine
The system operates as a network of agents, each responsible for a specific component: cooling towers, condenser pumps, chilled water pumps, and the chiller itself. A supervisor agent oversees all of them.
Each agent makes decisions by observing the environment (temperatures, flow rates, energy consumption) and adjusting equipment in response (like changing fan or pump speeds). If the result moves toward the goal (efficient cooling using less energy), it gets rewarded. If not, it gets penalized. That feedback loop helps it improve over time.
At the core is Q-Learning, a type of model-free RL algorithm that’s well-suited to systems like these, where there are many possible states and actions, and outcomes unfold over time.
Field test: DeJoule in action
To see how well it works, the team deployed DeJoule’s RL engine in a hospital in Thiruvananthapuram. The building had all the necessary instrumentation and cooling demands to provide a real-world testbed. To see how well it works, the team deployed DeJoule’s RL engine in a hospital in Thiruvananthapuram. The building had all the necessary instrumentation and cooling demands to provide a real-world testbed.
They ran a 30-day comparison:
- Before: The system was controlled using Joule Recipes' scripts (i.e., manual rules).
- After: The RL agents in DeJoule took over.
They didn’t just compare energy consumption averages; they normalized the comparison using a two-level binning method that accounts for ambient weather and cooling load. That way, the two timeframes were evaluated under equivalent operating conditions.
The results
DeJoule’s RL-driven control delivered real impact:
- Chiller plant efficiency improved by 5.33%
- Condenser pump usage dropped by 21.9%
- Chilled water pump output dropped by 6.7%
- Delta-T (a key efficiency indicator) improved by 6.6%
By intelligently adjusting flow rates and operating speeds, DeJoule allowed the chiller (the most energy-intensive component) to work less without compromising cooling. In effect, the system found the balance point between efficiency and performance and kept adjusting as conditions changed.
What we learned from deployments
After rolling out DeJoule with RL across 30+ hospitals, a few key lessons emerged:
- 1. High-quality sensor data is essential
RL can only be as good as the data it gets. Sensors must be accurate, well-maintained, and reliably connected.
- 2.User constraints can become bottlenecks
If users set unrealistic targets or too-narrow boundaries, the system wastes energy chasing unachievable goals. The team is now integrating deep learning to help set smarter targets that evolve with time.
- 3.Explainability is still a challenge
Like many ML systems, RL can feel like a black box. Understanding why the system made a particular decision isn’t always straightforward. DeJoule is working on better visualization tools to help users see the cause-and-effect relationships behind its decisions.
Conclusion
DeJoule isn’t just another BMS. It’s a full-stack platform designed to optimize every layer of building operations in real time. RL-driven chiller plant control is one piece of that, but a powerful one. It turns traditional HVAC management on its head—instead of fixed schedules and manual tweaks, buildings adapt minute by minute, learning and improving continuously.
With climate targets getting more ambitious and energy costs rising, this isn’t just a tech upgrade. It's a fundamental shift in how we run buildings. The results are already clear: better performance, lower emissions, less waste.