India collaboration ↗Proposed protocol, not a completed field trial. No plant modification authorised by this document.
Record owner permission, plant make/capacity from documentation, current condition, operating practice and competent installer review. Assess safe instrument location and calibration requirements. No electrical or gas-line modifications by an unqualified person. No compression, pressure experiment, deliberate overheating, feed shock or actuator command. Preserve existing safety protections. If the site is unsuitable, continue software demonstration separately and postpone field testing.
One row per observation with site_id, observation_id, local timestamp plus timezone, observer, feed type, feed mass_kg, water_added_L, ambient_temperature_C, process_temperature_C where available, gas_meter_reading, gas_unit, gas_reference_conditions, methane_fraction if measured, measurement_method, calibration_id, maintenance_event and quality_flag. Fields unavailable stay null with a reason. Do not infer methane composition from a screenshot or assumed fraction.
Keep a manual daily log first. Record a meter’s actual sampling cadence if one is safely installed. Do not interpolate daily observations to minutes and call them new measurements. Preserve raw intake with checksum, create a processing manifest and review corrections. Weather and plant temperature are distinct variables.
Check monotonic meter totals/reset events, duplicate timestamps, missingness, improbable units, outages and feed changes. An absence of alarms is not evidence of stable biology. Without measured gas flow, report record quality and operator usability only. Without VS/COD and compatible flow measurements, do not calculate a methane physics-violation rate as zero.
With sufficient target observations, pre-register forecast horizon and chronological splits before tuning. Fit preprocessing on training only. Start with persistence. Rolling origin uses pre-test periods; reserve a final period. Report sample counts, dates, MAE/RMSE, uncertainty and seed dispersion where meaningful. Farm data remain separate from DS-06. Do not pool incompatible targets or imply cross-plant validation from visual comparison.
Proposed target: at least 90% completion of agreed fields and a successful unassisted operator report after instruction. Forecast acceptance criteria must be chosen before the final evaluation, after feasibility assessment; no arbitrary guaranteed percentage improvement. Any failure or inadequate sample size remains a valid finding.
Retain immutable raw rows, processed checksums, Git/config versions, split dates, predictions, scenario IDs and operator approval notes. Store only consented project records. No public participant identity, farm location detail or private patent document upload. Offline UI must state data unavailable and the last observed timestamp. Physical control remains manual.