📖 About the Project
🔹 Inspiration
Buildings consume nearly one-third of global energy and contribute significantly to emissions. To support decarbonization, we focused on short-term load forecasting and anomaly detection in energy consumption, which can enable efficient energy management, better use of renewables, and reduced dependence on fossil fuels.
🔹 What We Learned
Through this project, we explored how Time Series Foundation Models (TSFMs) can capture general patterns across domains and how fine-tuning strategies improve performance on specific building types. We also deepened our understanding of evaluation metrics like MAE, RMSE, and MAPE, and how anomaly detection can complement forecasting models.
🔹 How We Built It
- Preprocessed building energy data with
pandasandNumPy(handling missing values, normalization, and feature engineering). - Established baseline models (ARIMA, Prophet, Random Forest Regressor).
- Fine-tuned pre-trained TSFMs
- Designed evaluation with:
$$
RMSE = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2}
$$
- Implemented anomaly detection using residual thresholds and statistical methods.
🔹 Challenges We Faced
- Handling irregular and missing time-series data.
- Balancing accuracy with limited training time in a hackathon setting.
- Selecting the best fine-tuning strategy among multiple approaches.
Built With
- hugging-face-transformers
- matplotlib
- numpy
- optuna
- pandas
- python
- pytorch
- scikit-learn
- statsmodels
