03Requirements
Combine hourly demand forecasts with weather, meter history and solar estimates so a user can explore building energy needs and renewable-energy scenarios.
OPENING THE SYSTEM
Loading the next page…
Loading the next page…
A full-stack application that combines building-demand forecasts, weather, meter history and solar estimates to explore energy-planning scenarios.
XGBoost forecasts building demand from weather, calendar features and meter history. FastAPI serves inference and solar-planning APIs; React makes forecasts, historical adjustments and planning estimates explorable. An optional PatchTST path runs when its checkpoint and sufficient history are available.
Predict building energy demand accurately and investigate AI-driven forecasting approaches for intelligent building energy management.
BDG-2 source dataset scale; not a training-row count
Validation covers four measured BDG-2 buildings. Monthly totals include a historical adjustment; the evaluation section explains the test window and its limits.
Combine hourly demand forecasts with weather, meter history and solar estimates so a user can explore building energy needs and renewable-energy scenarios.
BDG-2 contains approximately 26.5 million hourly observations, 1,512 buildings and 18 sites. Historical demand, weather and building metadata provide the inputs.
Keeping 720 hours of recursive predictions aligned with changing lag inputs, and making missing-data fallbacks and historical adjustments explicit.
BDG-2 supplies approximately 26.5 million hourly records spanning 1,512 buildings and 18 sites. Dataset scale describes the source corpus, not the number of rows used in a particular training run.
The active forecast model is BTX v3, an XGBoost model with temporal, weather, building and lag features. Optional PatchTST fusion requires a checkpoint and 336 continuous hours of meter history. The baseline remains available when those inputs are missing.
Built React and FastAPI layers around a 720-hour recursive forecasting path. Weather data and a climatology extension support the horizon. When recent meter history is available, monthly energy blends 20% raw model output with an 80% historical anchor, retaining the hourly forecast shape. Solar estimates use a physics-based model; regional tariffs are planning estimates.
The repository documents validation on 4 measured BDG-2 buildings for 1-30 August 2017, with 720 forecast hours preceded by 336 hours of history. It is an in-distribution check, not independent generalization. Monthly totals include the historical-anchor adjustment. One-hour forecast checks and monthly energy comparisons are reported separately in docs/validation.md.
Approximate BDG-2 dataset size; not a training-row count
BDG-2 across 18 sites
Recursive hourly forecast over a 30-day planning window
Continuous meter history required by the optional PatchTST path
Measured BDG-2 buildings; an in-distribution check, not independent generalization
A working application connects demand forecasts, measured history, weather and solar-planning estimates. Published validation distinguishes raw model predictions from adjusted monthly energy. Dataset size and forecast horizon describe engineering scope; they are not accuracy measures.
Keeping timestamps, units and lag indices aligned, making fallback paths visible, and separating model error from the effects of historical adjustments in long forecasts.
Data provenance, temporal structure and evaluation context matter as much as the model. A high score is useful only when its target and test conditions are clear.
Evaluate held-out buildings and future periods, compare raw and history-adjusted forecasts under the same conditions, and publish representative prediction traces.
AI-Based Energy Management in Building Operation: Integrating Renewable Energy