Short-term energy load forecasting

The building sector is responsible for nearly one-third of global energy consumption and carbon emissions, making it a priority for decarbonization initiatives. Effective short-term load forecasting is essential to optimize energy management, enable renewable integration, and reduce reliance on fossil fuel-based generation. However, conventional methods such as ARIMA, Prophet, or recurrent neural networks often require building-specific retraining, struggle with generalization, and lack robustness in detecting anomalies that signal inefficiencies or faults. This project proposes leveraging Time Series Foundation Models (TSFMs) to overcome these challenges. TSFMs are pre-trained on large, diverse time-series datasets and learn generalizable temporal patterns, enabling accurate forecasts for unseen buildings in a zero-shot setting and reducing the need for extensive retraining. Our approach focuses on two core tasks: • Short-term load forecasting – predicting energy demand over horizons ranging from minutes to days, using historical consumption data enriched with contextual features such as weather, occupancy, and calendar effects. • Anomaly detection – identifying significant deviations between predicted and actual consumption to detect operational anomalies, equipment failures, or abnormal energy usage. The methodology integrates data preprocessing, baseline modeling (classical and deep learning), and the application of both zero-shot and fine-tuned TSFM approaches. Performance will be evaluated using RMSE, MAE, and MAPE for forecasting accuracy, alongside Precision, Recall, and F1-score for anomaly detection.

Expected outcomes include enhanced forecast accuracy, scalable deployment across diverse building types, and early detection of energy anomalies, supporting proactive management. By enabling smarter demand response and greater renewable integration, the proposed solution contributes to sustainable, low-carbon building operations and advances the global clean energy transition.

Built With

• apache
• mlops
• python
• pytorch
• scikit-learn
• tsfm