A Hybrid Two-Stage Ensemble Framework for Detecting and Quantifying Energy Flexibility in Buildings: A Solution to the FlexTrack Challenge

Abstract: Accurate Measurement and Verification (M&V) of demand response (DR) is essential for integrating flexible building loads into the power grid. Addressing the FlexTrack 2025 Challenge, this paper details a hybrid two-stage ensemble framework to classify DR events and quantify their energy impact. The first stage uses a gradient boosting ensemble to predict the Demand Response Flag, identifying the building’s operational state. This classification is then fed as a key feature into a second-stage regression ensemble that estimates the Demand Response Capacity. The solution’s novelty lies in its hierarchical structure, which combines a general-purpose global model with specialized models trained on data-driven site archetypes for robust generalization. This methodology is underpinned by extensive feature engineering to capture complex temporal and weather-related dynamics. On the private test set, the solution achieved a Geometric-Mean Score of 0.618 for classification and a normalized Mean Absolute Error (nMAE) of 0.991 and normalized Root Mean Square Error (nRMSE) of 1.223 for regression. These results demonstrate the effectiveness of a decoupled, multi-model approach in tackling the complex challenge of DR baselining and provide a scalable framework for automated M&V systems.

---

About this Record: This Zenodo record provides a permanent, citable archive of the work described in the abstract. It contains:

• The full academic paper (Paper.pdf).
• The reproducible Python source code (flextrack_pipeline.py) and Jupyter Notebook.
• The code is also available on GitHub: DanGlChris - FlexTrack
• A Kaggle Notebook implementation is available here: FlexTrack 2025 Hybrid Ensemble Solution
• Data Access: The required data files are not included in this archive but can be downloaded directly from the FlexTrack Challenge on the AICrowd platform.