Optimise steam-to-carbon ratio to cut refinery emissions

Submitted through the COP31 Sustainable Transformation Awards · SKD Türkiye (WBCSD Global Network Partner)

The company operates a refinery in the energy sector, with more than 1,000 employees.

The refinery's hydrogen production unit makes hydrogen at 99.9 per cent purity from natural gas and supplies the hydrogen requirement of the other process units. Hydrogen is essential to the refinery's desulphurisation and conversion units, so the unit runs continuously and its operating cost is carried by every product that passes through those units.

The unit consumes refinery fuel gas and electricity while producing hydrogen, and it raises high-pressure steam in a waste heat boiler. Steam methane reforming, the technology used, is driven by the steam-to-carbon ratio: the amount of steam supplied for each unit of hydrocarbon feed. That ratio determines reformer conversion, the fuel gas burned in the reformer furnace, the steam recovered from the waste heat, and the natural gas charged as feed, and these effects pull in opposite directions.

In practice the ratio is often held at a conservative fixed value that protects the catalyst rather than one that minimises total cost, because the trade-off is difficult to calculate manually while process constraints move.

The objective was to lower the hydrogen production cost of the unit by optimising the steam-to-carbon ratio against total cost rather than against any single input. The year 2023 was taken as the baseline before implementation.

Location of the initiative: Türkiye

The initiative replaces a fixed operating setting with a calculated one, using a model built in-house from open-source technologies including Python and SQL rather than a licensed vendor optimisation package.

Model selection was empirical. Regression and tree-based models were both tested against unit data, and those performing above a defined threshold were taken through several validation stages before the final models were selected.

The models do not simply predict a set point. In each optimisation cycle the steam-to-carbon ratio is calculated iteratively within the given limits and the other process constraints, and the output of a total cost equation is produced for a range of different steam-to-carbon ratios. The operating point chosen is the one that minimises total cost across the whole balance, taking in refinery fuel gas saved, high-pressure steam recovered from waste heat, and the additional natural gas charged as feed, rather than the point that minimises any one of them.

Because the recommendation is derived from the cost equation rather than from a rule of thumb, the trade-off becomes explicit: the optimum accepts an increase in one input in exchange for larger reductions elsewhere.

Savings are calculated from flow meters on the unit, and the emissions effect is calculated in line with the ISO 14064 corporate carbon footprint standard. The application was commissioned in October 2023 and monitored over the following two-month period, and it remains in operation. Reporting was automated so that the result persists rather than depending on manual follow-up.

Figure 1: The STAR Refinery site at Aliağa, where the optimised hydrogen unit operates

Figure 2: The refinery complex housing the hydrogen production unit

Sustainability impact

Climate

The initiative targets Scope 1 emissions, the direct emissions from fuel combustion within the refinery, and specifically those of the hydrogen production unit.

The measured effect is a saving of 3,897 tonnes of CO2 a year. The emission calculation methodology follows the ISO 14064 corporate carbon footprint standard, savings are calculated from flow meters on the unit, 2023 is the baseline year before implementation, and the application was commissioned in October 2023 and monitored over the following two months.

The reduction comes from the energy balance rather than from any abatement equipment. Refinery gas consumption decreases and high-pressure steam production from waste heat increases. Hence, the equivalent amount of natural gas no longer has to be burned elsewhere to raise that steam. Against this, natural gas charged as feed for the steam-to-carbon optimisation has increased. The 3,897 tonnes of CO2 a year is the net result of that balance.

The measured boundary is the hydrogen production unit.

Social

The initiative was delivered by internal teams rather than by an external contractor, so the modelling capability stays in the refinery. Process, production, digital transformation and information technology teams worked together through the design, development and commissioning phases, in face-to-face and online sessions.

Building the model on open-source technologies means engineers can inspect, maintain and extend it, which makes the approach reusable on other units and reduces dependence on external specialists for future optimisation work.

Business impact

Benefits

• The direct benefit is a lower hydrogen production cost; the project delivered a 2 per cent saving in energy.

• Additional high-pressure steam raised from waste heat displaces steam that would otherwise be generated by burning fuel.

• Because the model was built on open-source technologies by internal teams, there is no software licence cost and no dependency on a vendor for changes.

• The company also expects a financial advantage from the reduction as emissions trading and carbon border mechanisms begin to apply to the sector, which is an anticipated benefit rather than a realised one.

Costs

The investment was engineering and analytics effort: process assessment, data extraction, model development and validation, commissioning and the automation of reporting, drawn from process, production, digital transformation and information technology teams.

The clearest trade-off is in the operating balance. Natural gas charged as feed increased, which is a real additional cost accepted because the refinery gas and steam gains are larger; the case therefore depends on the relative prices of natural gas, refinery fuel gas and steam, and a change in those prices moves the optimum.

Ongoing costs are model maintenance and data quality. The result depends on flow meter accuracy and on the models remaining representative as catalyst activity, feed composition and unit constraints change, so periodic revalidation is required.

Costs are contained by using open-source technologies rather than licensed optimisation software, by developing the model with internal teams, and by automating the reporting so that continuity does not consume analyst time.

Implementation cost and payback period are treated as internal company information and are not published.

Impact beyond sustainability and business

Co-benefits

• The method is reusable. The pattern of building a total cost equation for a unit, fitting models to historical operating data, and calculating the optimum operating point within process constraints applies to other refinery units where a single parameter drives several competing energy streams.

• Automating the reporting turns the optimisation into a monitored condition rather than a one-off project result, which is what allows the saving to persist after the project team has moved on.

• Because hydrogen units in refineries commonly use steam methane reforming from natural gas, the approach is applicable at refineries with similar technology, and the company notes that comparable projects were subsequently developed by other stakeholders in the sector.

Potential side-effects

• The optimisation increases one input while reducing others. Natural gas charged as feed rose, so a site evaluating the same change has to confirm that its own price and emission factors still make the balance favourable.

• Operating closer to a calculated optimum reduces the margin held by a conservative fixed steam-to-carbon ratio. Process limits and catalyst protection have to be encoded as hard constraints in the model, and operators need a route back to a safe fixed setting.

• The result depends on measurement. Savings are calculated from flow meters, so meter calibration becomes part of the emission and savings claim rather than a routine maintenance task.

• The model itself was not shared. Because it was built with internal resources and contains company-specific information, no work was carried out to disseminate its content, so what transfers to other operators is the method rather than the model.

Typical business profile

The model suits refineries and petrochemical sites operating hydrogen production units based on steam methane reforming from natural gas, which is the common configuration, and more generally any continuous process unit where one operating parameter drives several competing energy and feed streams.

It requires historical process data at sufficient resolution, flow measurement on the relevant streams, and a digital team able to build and validate models alongside process engineers who can define the constraints.

Delivery engages process engineering, production operations, digital transformation and information technology, with senior management approval before commissioning.

Approach

1. Select the unit by cost leverage: Identify a continuously operating unit whose production cost is driven by a single adjustable parameter, in this case the steam-to-carbon ratio of a hydrogen production unit.

2. Write the total cost equation before modelling: Express the cost of the unit as a function of every stream the parameter affects, here refinery fuel gas consumption, natural gas charged as feed and high-pressure steam raised from waste heat, so that the optimisation target is total cost rather than one input.

3. Establish measurement and a baseline: Confirm that flow meters cover the relevant streams, set a baseline year before implementation and fix the emission calculation method, in this case the ISO 14064 corporate carbon footprint standard.

4. Build the model with open-source technologies: Develop the optimisation model in-house using open-source tools such as Python and SQL, which keeps the licence cost at zero and the capability inside the engineering teams.

5. Test model families against each other: Try both regression and tree-based models, keep only those performing above a defined threshold, and take them through several validation stages before selecting the final models.

6. Encode the process constraints as hard limits: Calculate the ratio iteratively within the given limits and the other process constraints in each optimisation cycle, so that the recommended operating point is always feasible and safe.

7. Commission with production and monitor against the baseline: Obtain senior management approval, implement the recommended operating point with the production team, and monitor the result over a defined period, here the two months following commissioning in October 2023.

8. Automate the reporting for continuity: Work with the information technology team to automate online reporting of the saving, so that performance is tracked without manual effort and drift is visible.

Stakeholders involved

Project leads: The initiative was approved by senior management before commissioning. Delivery was led jointly by the technical service, production and digital teams rather than by a single owner, and the reporting was designed to be compatible with the site's digital infrastructure so that continuity is maintained after implementation.

Company functions: The process team assessed technical feasibility and supplied the process parameters and constraints. The digital teams built the model and ran the iterations that produced the optimum values. The production team took the actions required to apply the model in the field. The information technology team automated the online reporting used to track sustainability of the result over the long term. The teams worked in close contact, providing feedback through both face-to-face and online meetings.

Main providers: No external solution provider was engaged. The model was developed with internal resources using open-source technologies, which is why no licensed optimisation package appears in the delivery chain.

Other: Because the model was built internally and contains company-specific information, no dissemination work was carried out on its content. The company notes that comparable projects were subsequently developed by other stakeholders in the sector, so the approach has spread as a method rather than as a shared model.

Key parameters to consider

• The project was completed in 2023 and remains in operation. The baseline is the year 2023, before implementation; commissioning took place in October 2023 and monitoring covered the following two months.

• The measured boundary is the hydrogen production unit. Savings are calculated from flow meters, and emissions are calculated in line with the ISO 14064 corporate carbon footprint standard.

• The economics depend on the relative prices of natural gas, refinery fuel gas and steam, since the optimum accepts more of one to save more of the others.

• Replication is limited by data rather than by technology: the approach needs historical operating data at sufficient resolution and reliable flow measurement on each affected stream.

• The company obtains independent third-party external assurance for its reported information.

Implementation and operations tips

• Optimise the cost equation, not the parameter. A steam-to-carbon ratio chosen to minimise fuel gas alone would have missed the balance that produced the result, because the gain comes from accepting more feed gas in exchange for larger fuel and steam savings.

• Open-source tools are sufficient for this class of problem. Regression and tree-based models built with Python and SQL delivered a result that persists in operation, without licence cost or vendor dependency.

• Validate before trusting. Testing several model families and passing the survivors through several validation stages is what allowed the output to be used to set a live operating point on a continuously running unit.

• Automate the reporting on day one. A saving that has to be recalculated by hand will not be visible a year later, and the automated report is what proves the result has not drifted.

• Keep process safety limits outside the optimisation. The model searches within the limits it is given, so the constraints have to be defined by process engineers, not inferred from the data.