Cross-Sector Greenhouse Gas Scenario Model

Related work: this project's uncertainty and evidence-based approach follows the same research direction as trustlens-ai (calibrated, human-governed ML) and hospitality-sustainability-ai (real-world operational validation).

A small, transparent Python model for estimating greenhouse-gas emissions and testing reduction scenarios across electricity, heating, transport, waste, and industrial activity.

This portfolio project was created to demonstrate reproducible modelling, scenario design, uncertainty analysis, documentation, and version-control-ready research practice. It is a prototype, not a regulatory carbon-accounting tool.

Research agenda: Transparent Cross-Sector Emissions Scenarios Under Parameter Uncertainty

At a glance

Research element	Current implementation
System	Electricity, heating, transport, waste and industrial modules
Core method	Auditable activity × emission factor calculations
Intervention analysis	Editable baseline, balanced-transition and rapid-decarbonisation scenarios
Uncertainty	One-at-a-time sensitivity and reproducible Monte Carlo simulation
Applied case	Hospitality water-energy-emissions scenario
Scientific ML	Reproducible data-only vs physics-informed neural benchmark
Real-data validation	Held-out EPA eGRID natural-gas plant comparison
Interpretation	Conditional scenario estimates, not measurements or forecasts

Research question

How do different sector-level interventions change total annual greenhouse-gas emissions, and how sensitive are the results to uncertainty in activity data and emission factors?

Research contribution

The project's research contribution is a compact, provenance-first workflow for testing whether mitigation priorities remain stable when activity data, emission factors, uncertainty assumptions and system boundaries change. The goal is not to produce the largest headline reduction. It is to show which conclusions are reproducible, which inputs dominate uncertainty and when an apparently clear intervention ranking is too unstable to defend.

The next study phase adds a versioned factor registry, correlated uncertainty, rank-reversal analysis and reconciliation against an independently prepared inventory. See the research agenda for the proposed questions, hypotheses and publication boundaries.

Method

For each activity record:

emissions (kg CO2e) = activity × emission factor

A scenario changes activity, the emission factor, or both. Independent relative uncertainties are combined using root-sum-of-squares propagation.

Included scenarios

• baseline: no intervention
• balanced_transition: lower electricity demand, grid decarbonisation, heating efficiency, transport electrification, waste reduction, and industrial efficiency
• rapid_decarbonisation: more ambitious reductions across the same sectors

Quick start

python -m venv .venv
.venv\Scripts\activate
pip install -e .
ghg-model --activities data/sample_activities.csv --scenarios data/scenarios.json --output outputs/results.csv
ghg-chart
python -m unittest discover -s tests -v

On macOS/Linux, activate the environment with source .venv/bin/activate.

Example output

Hospitality water-use case study

The repository also includes a small case study motivated by firsthand observation of restaurant operations: taps may remain running during repeated utensil-rinsing tasks. This is treated as a research question, not as a claim about a named employer or a formally measured site.

The demonstration asks: How much water and associated water-system emissions could be avoided by reducing unnecessary tap-running time while preserving food-safety and service requirements?

ghg-water-case

The default example uses transparent, editable assumptions:

• tap flow: 6 litres per minute (illustrative; measure before real use)
• running time: 7.5 hours per operating day (midpoint of a 7-8 hour observation)
• operating days: 300 per year (illustrative)
• avoidable-use reduction: 75% (scenario assumption, not a forecast)
• hot-water share: 50% (illustrative)
• inlet/outlet temperature: 12/45 degrees C (illustrative)
• heater efficiency: 90% and heating factor: 0.18231 kg CO2e/kWh

It applies UK Government 2026 water-supply, water-treatment and natural-gas factors. Water-heating demand is calculated from water mass, specific heat capacity, temperature rise and heater efficiency. Water temperature, hot-water share and heater technology were not measured, so the defaults are explicitly illustrative and the CLI allows every value to be replaced. Water-system and heating emissions are reported separately to keep the system boundary auditable.

One-at-a-time sensitivity analysis

The sensitivity command varies one input at a time around the base case, recalculates potential total emissions savings and ranks parameters by their largest percentage effect on the output:

ghg-sensitivity --variation-pct 10

The default ±10% range is a local diagnostic, not a confidence interval. It answers which inputs the model responds to most strongly near the selected base case. It does not represent correlations, probability distributions or uncertainty in the observed 7-8 hour running-time statement. Those questions require a separate uncertainty analysis.

Monte Carlo uncertainty analysis

The Monte Carlo command samples transparent triangular ranges for flow, running time, operating days, avoidable use, hot-water share, temperatures, heater efficiency and the heating emission factor:

ghg-monte-carlo --samples 10000 --seed 2026

The fixed seed makes results reproducible. Inputs are currently sampled independently, which is a simplifying assumption because real operating variables may be correlated. The reported P5-P95 interval is conditional on the selected ranges and distribution shapes; it is not a measured confidence interval and should be updated when field data become available.

Least-cost abatement pathway optimisation

The optimisation command finds the lowest-cost combination of candidate measures that meets a stated emissions-reduction target, treating each measure as an indivisible (0/1) adoption decision solved via mixed-integer linear programming. A companion rank-reversal analysis perturbs every measure's cost and abatement within its stated uncertainty and reports how often each measure survives in the optimal set — distinguishing genuinely robust choices from ones that only look optimal under one set of point estimates.

ghg-optimize --options data/abatement_options.json --target-kgco2e 2500 --rank-reversal

On the illustrative eight-measure dataset, a 2,500 kg CO2e/year target is met by five measures at roughly £20/year net cost. Under 1,000 Monte Carlo perturbations, the two net-saving measures are selected in 100% of trials, while the single largest-abatement measure in the dataset is selected in under 2% of trials. This is a demonstration of the method's ability to separate robust choices from point-estimate artefacts, not a real efficiency finding for any venue. All costs and abatement figures are illustrative placeholders. See docs/optimization_pathway.md for the complete method, worked example and limitations.

Scientific-machine-learning foundation experiment

The repository includes a compact neural-surrogate benchmark that compares a data-only objective with a physics-informed objective containing the known emissions = activity × emission factor relationship. It evaluates both held-out interpolation and deliberate extrapolation beyond the synthetic training range:

ghg-sciml

This is an educational foundation experiment on synthetic data, not industrial validation or evidence of expertise with scientific foundation models. The complete design and progression path are documented in docs/scientific_ml_benchmark.md.

The locked default run reduced interpolation RMSE by 35.8% and extrapolation RMSE by 12.2% relative to the identical data-only network. Extrapolation error remained high, which is reported as a negative result and a reason to study stronger architectures, uncertainty methods and genuine scientific datasets.

First real-data validation

The first real-data benchmark compares estimates against reported annual CO2 from held-out natural-gas plants in an official EPA eGRID-derived 2018 layer. It excludes target-derived emission-rate fields and uses the EPA GHG Emission Factors Hub value of 53.06 kg CO2/MMBtu as an external physical constraint:

ghg-real-validation

This is an unseen-plant test within one historical US data year. It is not temporal, contemporary, European or cross-sector validation. The locked design, provenance, leakage controls and interpretation boundaries are documented in docs/real_data_validation.md.

On the locked 320-plant holdout, the physics-guided residual model reduced median absolute percentage error from 1.88% to 1.43% relative to the data-only ridge model. Its RMSE was 0.06% worse, so the project reports the improvement in typical relative error without claiming that the hybrid wins every metric.

Data and provenance

data/sample_activities.csv contains a synthetic demonstration dataset. Its emission factors are illustrative and must not be used for formal reporting. The model keeps factor year, source, unit, and uncertainty alongside every record so authoritative factors can be substituted without changing the code.

For a real study, use the latest UK Government greenhouse-gas conversion-factor flat file or another jurisdiction-appropriate inventory source, document all mapping decisions, and retain the original source version. Relevant methodological foundations include:

• UK Government greenhouse-gas reporting conversion factors: https://www.gov.uk/government/collections/government-conversion-factors-for-company-reporting
• UK Government conversion factors 2026: https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2026
• UK Government 2026 methodology report: https://assets.publishing.service.gov.uk/media/6a2940543b15d05a7ce3202e/2026-GHG-conversion-factors-methodology-report.pdf
• 2006 IPCC Guidelines for National Greenhouse Gas Inventories: https://www.ipcc-nggip.iges.or.jp/public/2006gl/

Repository structure

data/                  Synthetic activities and scenario assumptions
src/ghg_model/         Model and command-line interface
tests/                 Unit tests
outputs/               Generated results (created when the model runs)

Limitations and next steps

• Replace illustrative factors with an automated, versioned import from an authoritative source.
• Add direct and indirect emissions with explicit Scope 1, 2, and 3 treatment.
• Add spatial and hourly resolution for integration with an energy-system model.
• Model correlations and probability distributions using Monte Carlo simulation.
• Validate sector mappings and scenario assumptions with domain experts.
• Test whether intervention rankings reverse across plausible input distributions.
• Reconcile one complete case against an independently prepared inventory.

Ethical and reproducibility note

Results are only as reliable as their boundaries, data, factors, and assumptions. Scenario outputs should be reported with uncertainty and should not be presented as forecasts.