The national context for green hydrogen research
The UAE National Hydrogen Strategy sets production targets of 1.4 million tonnes per year by 2031, comprising 1 million tonnes of green and 0.4 million tonnes of blue hydrogen, rising to 7.5 million tonnes by 2040 and 15 million tonnes by 2050. It also plans hydrogen oases and a national research and development centre for hydrogen technologies (UAE Government, n.d.). University research is organised around similar themes: Khalifa University's Research and Innovation Center on CO2 and Hydrogen (RICH) lists hydrogen production and storage, studied through combined modelling and experiment, and techno-economic and life cycle assessment among its themes (Khalifa University, n.d.).
For a doctoral candidate this context is useful but double-edged. It makes a question about cost, scale-up or forecasting relevant to policy. It also tempts candidates to collect large data sets before they have a question the data can answer. Most weak analysis chapters can be traced to that reversal.
How UAE researchers have analysed green hydrogen
The studies below were written by authors at UAE institutions, with affiliations checked on the publisher page. They cover the main analytical families a PhD is likely to use.
Techno-economic modelling of a building system
Urs et al. (2023), at Khalifa University, modelled a grid-connected microgrid for a medium-sized office building in Abu Dhabi, using solar photovoltaics, an electrolyser, a hydrogen tank and a fuel cell. Their inputs show where such data come from: a load profile from an Abu Dhabi Urban Planning Council benchmarking project, solar irradiance from the NASA POWER database, and component costs from the literature. Optimising the system in HOMER, they reported a levelised cost of hydrogen of about US$3.12 per kg, a renewable fraction of 78% and three-quarters of the load met without the grid. Their sensitivity analysis showed that capital costs and discount rates strongly affected the economics, and they noted a practical constraint: net-metering rules capped photovoltaic capacity.
Modelling national adoption
Gandhi et al. (2022), at the Dubai Electricity and Water Authority research centre, built an integrated adoption model for UAE industry. They projected that green hydrogen could become economically viable in UAE industries between 2032 and 2038 at prices of roughly US$0.95 to 1.35 per kg, depending on electrolyser and solar cost projections, and cautioned that meeting net-zero goals might need further measures such as carbon pricing if cost reductions slowed. The analytic problem here is not data noise but uncertainty about the future, handled through scenarios.
System-level and energy-flow analysis
Zaiter et al. (2023), mainly at Khalifa University, analysed national energy-flow diagrams, emission trends and the social cost of carbon to ask where hydrogen could help decarbonise industry. They concluded that decarbonising industry through renewable electricity alone faces obstacles, and that domestic solar-powered green hydrogen offers a route for hard-to-abate sectors.
Forecasting the solar input
Almarzooqi et al. (2024), also at Khalifa University, developed a hybrid kernel ridge regression model for short- and medium-term solar irradiance forecasting, compared it with persistence and regression baselines, and validated it with a moving-window scheme suited to time series. Two points matter for a PhD. They identified cloudy, variable conditions as the hardest case. And their data came from the US National Solar Radiation Database for two American sites, which shows that even UAE-based teams may depend on public data from elsewhere when local measurements are not available.
Optimising an electrolysis experiment
Rezk et al. (2023), a team including A. G. Olabi of the University of Sharjah, modelled hydrogen production from water electrolysis with an adaptive neuro-fuzzy inference system and optimised operating conditions with a metaheuristic algorithm. The model was fitted to 17 experimental runs with a 70:30 training and test split, and reported a near-perfect training fit. For a doctoral reader the method is instructive, and so is the sample size: with so few runs, independent confirmation experiments and uncertainty estimates are essential before optimised settings are trusted.
| Study | Analysis | Data | Main problem and handling |
|---|---|---|---|
| Urs et al. (2023) | Techno-economic optimisation (HOMER) | Local load profile; NASA POWER irradiance; literature costs | Cost uncertainty; sensitivity to capital cost and discount rate |
| Gandhi et al. (2022) | Adoption modelling | Cost and price projections | Future uncertainty; scenarios |
| Zaiter et al. (2023) | Energy-flow and emissions analysis | National energy and emissions data | System boundaries; framework for hard-to-abate sectors |
| Almarzooqi et al. (2024) | Irradiance forecasting | US public irradiance data | Cloudy conditions; moving-window validation |
| Rezk et al. (2023) | ANFIS model and optimisation | 17 experimental runs | Small sample; train and test split |
Studies of the UAE by authors elsewhere help calibrate these results. Joubi et al. (2022), at the University of Tsukuba, used the System Advisor Model to compare solar thermal and photovoltaic routes in the UAE and found levelised hydrogen costs of about US$7.85 to 10.28 per kg under their assumptions, including a 6% discount rate. Meighan et al. (2025), at Newcastle University, judged the 2031 target hard to reach without more domestic research, regulatory alignment and infrastructure, and listed water scarcity and reliance on imported electrolysers among the threats. The spread between these cost estimates and those above is itself a finding: it reflects different scales, technologies and assumptions, which is why a thesis must state its own.
Start from a question the data can answer
Each study above began with a specific question: what does hydrogen storage cost in one building type, when does adoption become economic, how accurately can irradiance be forecast. A doctoral question should be equally specific about the system, the outcome and the comparison. "How can green hydrogen be optimised in the UAE?" is a topic. "How does the levelised cost of hydrogen from a photovoltaic-alkaline electrolyser system in Abu Dhabi change with electrolyser capital cost, capacity factor and discount rate between 2026 and 2035?" is a question that fixes the data, variables and analysis.
- Name the system boundary: production only, or storage, transport and use as well.
- Name the outcome: levelised cost, emissions, reliability or forecast error.
- Name the drivers you will vary and the ones you will hold fixed, with sources.
- Name the comparison: another technology, location, year or policy scenario.
Originality follows from the same discipline. Applying an established model to a new data set is rarely enough on its own for a doctorate. The contribution can come from a new comparison (for example, a technology or location the UAE literature has not yet costed), from better local data (measured load or irradiance where earlier studies relied on databases or literature values), from a more rigorous treatment of uncertainty, or from linking two analyses that are usually done separately, such as forecasting the solar input and costing the hydrogen output in one model. The evidence review above is a practical way to find the gap: list what each study assumed, measured and left out, and look for the omission that matters most to your question.
Match the method to the data
Techno-economic studies need a transparent cost model. The levelised cost of hydrogen divides the discounted lifetime costs (capital, operation and maintenance, electricity, water and stack replacement) by the discounted lifetime hydrogen output. Report every input with its source and year, and present sensitivity or Monte Carlo analysis for the inputs with the widest plausible ranges, as Urs et al. (2023) did for capital costs and discount rate.
Forecasting studies need validation that respects time. Random k-fold cross-validation can leak information from the future into training; rolling-origin or blocked schemes are safer, and Bergmeir and Benítez (2012) discuss when standard cross-validation can and cannot be used for time-series predictors. Always report a naive baseline such as persistence, as Almarzooqi et al. (2024) did, so that gains are measured against something meaningful.
Experimental modelling needs enough independent runs to estimate error. With fewer than about 20 runs, prefer simpler models, report prediction intervals and confirm optimised settings with new experiments rather than relying on a held-out handful of points.
Worked example: how two assumptions move the levelised cost
The numbers below are illustrative and hypothetical, chosen to show the arithmetic, not taken from any study. Consider a 1 MW electrolyser costing US$1,000 per kW installed, with a 20-year life, using 55 kWh of electricity per kg of hydrogen bought at US$0.03 per kWh.
- At a 25% capacity factor the plant runs 2,190 full-load hours a year and produces about 39,800 kg of hydrogen.
- The capital recovery factor is about 0.087 at a 6% discount rate and 0.117 at 10%, so annualised capital is about US$87,200 or US$117,500.
- Capital therefore contributes about US$2.19 per kg at 6% and US$2.95 per kg at 10%; electricity adds US$1.65 per kg in both cases.
- Raising the capacity factor to 35% cuts the capital share at 6% to about US$1.56 per kg.
- Operation, maintenance, water and stack replacement are still to be added, so the full cost is higher.
Two assumptions alone move the result by more than a dollar per kg. That is why a thesis should present a tornado chart or similar sensitivity summary and explain which inputs most need better local data.
What to report
| Element | What to include |
|---|---|
| Question and boundary | System, outcome, comparison and time horizon |
| Data provenance | Source, location, period and resolution of solar, load and cost data |
| Model | Software and version, equations, fixed and varied inputs |
| Validation | Baselines, time-aware validation, independent test data |
| Uncertainty | Sensitivity ranges, scenarios or prediction intervals |
| Policy link | How results relate to national targets, stated without overclaiming |
Common mistakes
- Collecting large data sets before defining the question and the outcome.
- Reporting a single levelised cost with no sensitivity analysis.
- Using irradiance or cost data from another country without saying so or testing the effect.
- Validating forecasting models with shuffled cross-validation.
- Trusting near-perfect fits on very small experimental data sets.
- Claiming a study shows how the national target will be met when it models one site or technology.
Key takeaways
- UAE-based studies show that levelised costs depend heavily on capital cost and discount rate, so report ranges (Urs et al., 2023; Joubi et al., 2022).
- Future uncertainty is best handled with explicit scenarios (Gandhi et al., 2022).
- Forecasting models need time-aware validation and naive baselines (Almarzooqi et al., 2024; Bergmeir & Benítez, 2012).
- Small experimental data sets call for simple models and confirmation runs (Rezk et al., 2023).
- Link findings to national targets carefully and state the limits of a single-site study (UAE Government, n.d.; Meighan et al., 2025).
- System-level analysis helps show where hydrogen adds value beyond renewable electricity (Zaiter et al., 2023).
