GFGFIS Evidence PortalChatake Innoworks · Version 3
Data · chronology · epochs · decision

Measured-methane training is visible, replayable and judged against simpler baselines.

The current temporal study uses evaluator-suggested DS-05 reactor observations for measured methane. DS-06 contributes 500,400 full-scale SCADA rows for industrial replay and a future total-biogas benchmark; the two targets remain separate.

Evaluator-suggested public evidence

138
DS-05 reactor records

Dated March–July observations retain reactor identity and measured process fields.

86
Measured-methane targets

Methane is calculated only where total biogas and methane fraction coexist in the source.

500,400
DS-06 industrial rows

One-minute full-scale SCADA supports observed plant replay. Its Biogas channel is total biogas in CFM—not methane.

Target statement: DS-05 supports the present measured-methane comparison. DS-06 is an industrial operating-history study and replay foundation. No DS-06 neural epoch is claimed until its published cleaning procedure and leakage-safe split are reproduced.

Leakage-safe experiment design

Immutable raw archive
Canonical reactor schema
Chronological pre-test folds
Fold-local preprocessing
Locked final period

What training is allowed to see

Only observations earlier than the current validation block. Imputation medians, scaling statistics, lag features and sequence windows are fitted or constructed within the training chronology.

What training cannot see

No random shuffle, no validation-fitted preprocessing, no test-period model selection, no future target in lag features and no cross-reactor sequence mixing.

Open chronological split manifest Open run manifest

Epoch-by-epoch measured-methane training

This figure is regenerated from run TE-DS05-20260724T212528Z. Every line is a retained training session identified by rolling fold and seed. Click to enlarge it.

Actual DS-05 LSTM training loss across rolling folds and seeds

Actual DS-05 LSTM training-loss histories for the five-observation window. Training loss explains optimisation behaviour; architecture selection still uses rolling-origin validation metrics.

Configuration

  • Measured-methane target
  • Sequence window 5
  • Maximum 60 epochs
  • Seeds 17, 42 and 73
  • No random time-series shuffle
  • Past-only features and chronological folds

Open actual outputs

Every retained epoch
Rolling-fold metrics
Untouched final-test metrics
Run configuration and replay command

Why the baseline matters

DS-05 persistence, XGBoost and LSTM rolling-fold robustness

Across rolling-origin DS-05 validation, no temporal architecture receives promotion from one favourable score.

DS-05 untouched final-period comparison

On the untouched final period, XGBoost has the lowest mean error across repeated seeds; the LSTM loses to both required baselines.

Architecture decision

XGBoost is the final-test leader, while rolling-origin average RMSE remains close to persistence. The compact LSTM stays a research challenger—not the production champion. Longer memory or Transformers are not justified by this run.

What the 500,400-row study adds next

Industrial chronology

Nearly one year of one-minute temperatures, feed, level and total-biogas signals exposes operating transitions absent from the smaller reactor study.

Benchmark boundary

The next experiment must reproduce the published 491,761-row cleaned set before fitting persistence, XGBoost or LSTM models.

Plant replay

The indexed raw chronology already drives the 3D plant as observed SCADA. Model output and simulator output remain separately labelled.