Persistence
Remains essential because a temporal neural model must beat the previous observation under leakage-safe validation.
Dataset provenance register
| ID | Source | GFIS use | Scientific boundary |
|---|---|---|---|
| DS-05 | Mendeley Data CC BY 4.0 | 138 March–July reactor records; 86 measured-methane targets; chronological persistence/XGBoost/LSTM comparison. | Small research-reactor series; missing laboratory variables remain missing. |
| DS-06 | University of Iowa ODC-By 1.0 | 500,400 one-minute full-scale SCADA records; industrial replay, data quality and future total-biogas benchmark. | Biogas is total biogas in CFM—not methane. No DS-06 neural epoch is claimed yet. |
| DS-03 | Mendeley farm-scale monitoring | Earlier total-biogas and weather/temperature-regime experiment retained as comparative evidence. | Does not independently validate continuous methane because methane fraction is substantially missing. |
Current experiment decision
XGBoost
Lowest repeated-seed final-period DS-05 error; rolling-origin robustness remains the primary decision criterion.
LSTM
Actual 60-epoch histories are retained, but the present LSTM loses to required baselines on the untouched period.
Research-paper dossier
Paper A · temporal methane intelligence
Leakage-safe public-data comparison, limitations, drift and reproducible experiment memory.
Paper B · physics-guided plant twin
Mass/physics boundaries, VFA/ALK calibration pathway, scenario memory and operator evidence.
Industry benchmark programme
Continuous methane analyser data, laboratory VFA/ALK, VS/BMP physics-loss ablation and plant-specific calibration remain required.