03Architecture
01 / Building data
Hourly demand, weather observations and building metadata from BDG-2: approximately 26.5 million rows, 1,512 buildings and 18 sites.
OPENING THE SYSTEM
Loading the next page…
Loading the next page…
From millions of building observations to a clearer picture of tomorrow’s energy demand.
A forecasting pipeline combining temporal feature engineering, XGBoost, transformer development and model fusion on the BDG-2 building dataset.
Predict building energy demand accurately and investigate AI-driven forecasting approaches for intelligent building energy management.
Reported research results · target units not supplied
XGBoost uses the raw electricity target. Fusion R² ≈ 0.960 comes from a separate aligned log-target evaluation; the two scores should not be ranked directly.
Hourly demand, weather observations and building metadata from BDG-2: approximately 26.5 million rows, 1,512 buildings and 18 sites.
AI-Based Energy Management in Building Operation: Integrating Renewable Energy
Hourly time series span different buildings and sites. Evaluation needs to distinguish raw and transformed targets, align model outputs and account for the accumulation of error during recursive forecasting.
BDG-2 provides hourly observations, historical energy demand, weather information and building metadata. The stated dataset scale is approximate.
XGBoost provides a feature-driven approach, while Transformer / PatchTST development explores temporal representations. Fusion requires an aligned evaluation set and a consistent target transformation.
The work covers temporal feature construction, lag-based predictors, transformer iterations, model fusion, recursive forecasting, external validation and solar generation modelling.
Tree-based prediction and transformer-based temporal modelling are investigated as complementary forecasting approaches. Exact architecture hyperparameters are not yet published here.
Reported XGBoost raw-electricity test results: R² ≈ 0.983, RMSE ≈ 31–32 and MAE ≈ 10.7. Fusion R² ≈ 0.960 belongs to a separate aligned log-target evaluation; these R² values are not a head-to-head ranking.
XGBoost · raw electricity target
Approximate BDG-2 dataset size
Across 18 sites
Raw electricity target; units not supplied
Raw electricity target; units not supplied
Separate aligned log-target evaluation
The supplied metrics show strong fit on the reported XGBoost test evaluation. They do not establish deployment performance or energy savings. Prediction traces and feature-importance exports have not been supplied.
Key technical challenges include temporal feature construction at dataset scale, transformer iteration, target alignment for fusion and the compounding uncertainty of recursive predictions.
A model score needs its target, split and evaluation context. Comparing raw-target and log-target results without that context would hide the actual engineering question.
Publish reproducible evaluation artefacts, temporal prediction exports and feature-importance data. Extend external validation and document recursive forecast behaviour.
The repository link has not been published. Ask about the implementation