OCCUPANCY FORECASTING · SAVED-OUTPUT DEMO
When should a forecast
become a recommendation?
A high score for an empty building period is only a starting point. See how the score threshold and required duration affect which periods a model recommends.
Try the recommendation rule
Start with the paper's validation-selected settings, then change a control to explore the same historical records. No model training or installation is needed.
Loading the saved records…
The demo data could not be loaded. Check your connection and try again, or use the notebook below.
The load proxy is an overlap measure, not measured energy savings or verified controllable capacity. Zero recorded conflicts does not guarantee future performance or physical absence.
LOOK AT ONE HORIZON
Scores, labels, and recommendations
The default example has the largest load-proxy overlap under the reported hybrid policy. A shared link may select another date. The metrics above always cover all 43 horizons.
Loading the example…
Each interval is 15 minutes. Local Pacific labels run from 00:15 to the next 00:15. The completed 00:00 input bin is available at 00:15.
Read the example as a table
| Interval start | Empty score | Camera label | Recommendation | Load proxy (kWh) |
|---|
Compare the reported policies
Each model below uses its own threshold selected on validation data, with a one-hour minimum window. This table stays fixed when you explore the controls.
| Model | Threshold | Windows | Intervals | Conflicts | Load proxy (kWh) |
|---|---|---|---|---|---|
| Selected hybrid | 0.875 | 14 | 259 | 0 | 490.1 |
| Historical schedule | 0.950 | 12 | 66 | 0 | 94.6 |
| LightGBM | 0.950 | 19 | 265 | 11 | 493.9 |
The manuscript's AUPRC comparison uses 388,032 overlapping forecast rows, a different scope from these 4,128 unique policy intervals. The hybrid's small AUPRC advantage does not establish statistical superiority.
Inspect it. Reproduce it. Tell us what you find.
The notebook recomputes this comparison from the same compact records. A reproduction report helps other readers understand which environments work and where the instructions need improvement.
Download the compact source records (JSON) · Contribution guide · Read the manuscript
Data provenance and interpretation
The demo uses the cleaned LBNL Building 59 release and the repository's saved test predictions. Its hybrid uses the fixed weights 0.15 historical schedule, 0.60 LightGBM, and 0.25 compact Transformer. It does not select new weights or fit a controller. The reported thresholds met a 10% empirical validation conflict cutoff, which is not a guarantee about future risk.
Moving the controls explores an already inspected test period. It cannot establish a better held-out policy or new confirmed results. Source timestamp provenance, upstream imputation, corrected deep-model initialization, and evaluation on fresh data remain necessary for a prospective study.
Dataset provenance · Full limitations