ZeroGrid

ZeroGrid – Towards Net-Zero Buildings

Inspiration

Our starting point was a simple but powerful question: what if buildings could learn how much energy they’ll need before they actually use it?
Since buildings are responsible for about a third of the world’s energy consumption, we knew even small improvements in efficiency could have a massive impact. The push for net-zero buildings—spaces that generate as much renewable energy as they consume—motivated us to design something practical, impactful, and scalable.

What it does

ZeroGrid predicts how much energy a building will use in the near future and flags unusual patterns that could point to waste or faults.

If usage is higher than expected, it could mean equipment isn’t working properly.
If forecasts are accurate, building managers can adjust consumption or balance renewable energy better.

In short, ZeroGrid helps buildings waste less, save more, and move closer to net-zero goals.

How we built it

We began by collecting building energy data along with weather information and occupancy trends. After cleaning and normalizing the data, we experimented with both classic forecasting methods like ARIMA and modern deep learning models.

The heart of ZeroGrid lies in Time Series Foundation Models (TSFMs), which are pre-trained on a wide range of time-based data. By fine-tuning them on building energy data, we were able to get strong predictions without starting from scratch.

For anomaly detection, we used a simple but effective approach: compare actual usage with predicted values, and if the difference
[ \text{Error} = |y_{actual} - y_{predicted}| ]
crosses a certain threshold, we flag it as an anomaly. Finally, we built dashboards to make the results easy to interpret.

Challenges we ran into

Messy data: Some datasets had missing or inconsistent values, which slowed down preprocessing.
Balancing generalization and accuracy: We wanted our model to work across many building types, but that sometimes meant losing accuracy on a specific one.
Setting anomaly thresholds: Too strict, and you get false alarms. Too loose, and you miss real issues.
Integration: Forecasting and anomaly detection had to flow seamlessly for end-users.

Accomplishments that we're proud of

We managed to get accurate forecasts using TSFMs with only light fine-tuning.
We proved that anomaly detection can be integrated directly with forecasting instead of being a separate system.
Our dashboard made complex AI outputs understandable for facility managers, not just data scientists.
Most importantly, we created a project that aligns with real-world sustainability goals.

What we learned

The importance of data preparation—good predictions start with clean inputs.
Foundation models are extremely powerful when paired with domain-specific knowledge.
Forecasting isn’t just about numbers—it’s about enabling smarter decisions in the real world.
Even modest improvements in accuracy can lead to big environmental and financial savings.

What's next for ZeroGrid

We don’t want ZeroGrid to stay a prototype. Next steps include:

Expanding to larger datasets and more diverse building types.
Deploying in real-time with IoT-enabled smart meters.
Experimenting with adaptive anomaly detection that evolves with building behavior.
Scaling beyond individual buildings to model entire districts or cities.
Partnering with energy providers to turn ZeroGrid into a real sustainability tool.

ZeroGrid – Smarter forecasts for a net-zero future.