Inspiration

We’re building TSFM-EDIFice to prove that foundation models can forecast and detect energy anomalies, helping buildings cut emissions faster than ever. Excited to tackle decarbonization, we use TSFMs for zero-shot energy forecasting and anomaly detection making smart, green buildings a reality.

What it does

Zero-Shot Building Energy Forecasting with TSFM provides accurate short-term load forecasts and real-time anomaly detection for commercial and residential buildings. It leverages pre-trained Time Series Foundation Models (TSFMs) to deliver zero/few-shot predictions without the need for retraining, enabling fast deployment, smarter HVAC scheduling, and better renewable energy integration.

How we built it

We started with state-of-the-art TSFMs trained on diverse time series data. We designed a pipeline that takes building metadata, weather, and historical consumption as inputs, then generates 15-min to 48-hour forecasts. For anomaly detection, we combined residual thresholding, unsupervised TSFM embeddings, and statistical change-point methods. We wrapped this in reproducible notebooks and built a simple dashboard to visualize forecasts, anomalies, and actionable insights.

Challenges we ran into

Adapting TSFMs to building energy data with limited labeled anomalies.

Ensuring zero-shot forecasts remain robust across diverse building types and weather conditions.

Balancing accuracy with lightweight inference for potential on-edge deployment.

Designing a dashboard that conveys complex model outputs in a simple way for operators.

Accomplishments that we're proud of

Demonstrated strong zero-shot performance on public datasets like ASHRAE and REFIT.

Built a reproducible, modular pipeline that others can adopt quickly. Designed an anomaly detection layer that captures both sudden faults and subtle drifts.

Created a dashboard prototype that turns forecasts into actionable building insights.

What we learned

We learned how powerful foundation models are for transfer learning in energy applications. Even without retraining, TSFMs generalize well, and with light fine-tuning, they can rival or outperform traditional site-specific models. We also realized the importance of combining ML with domain knowledge (e.g., HVAC schedules, occupancy patterns) to improve trust and usability.

What's next for Zero-Shot Building Energy Forecasting with TSFM

Test on live building data and collaborate with facility managers.

Add adaptive fine-tuning modules for continuous improvement.

Expand the dashboard with predictive control suggestions (e.g., HVAC pre-cooling, battery scheduling).

Built With

  • and
  • and-feature-engineering-for-weather/calendar/metadata-inputs.-evaluation-&-benchmarking-?-public-datasets-like-ashrae-great-energy-predictor
  • and-uci-building-energy-datasets.-visualization-&-dashboard-?-streamlit/plotly-dash-for-interactive-dashboards;-matplotlib/seaborn-for-static-plots.-deployment-?-lightweight-docker-container-for-portability
  • edge
  • fine-tuning)-and-hugging-face-libraries-(for-time-series-foundation-model-integration).-data-handling-&-processing-?-pandas
  • inference
  • numpy
  • potential
  • pre-trained-time-series-foundation-models-(tsfms)-?-transformer-based-architectures-trained-on-diverse-temporal-datasets.-python-?-core-implementation-language.-deep-learning-frameworks-?-pytorch-(for-model-inference
  • quantization
  • refit
  • scikit-learn
  • supported
  • uk-dale
  • via
  • with
Share this project:

Updates

Submission history