← Back to the map

Methodology & sources

How the Green Hydrogen Cost Atlas turns a click on the map into a levelized cost of hydrogen (LCOH), and where every number comes from.

1. Plant architecture modeled

An off-grid, dedicated ("islanded") plant: a solar PV park, wind farm, or co-located mix whose only customer is a co-located electrolyzer. There is no grid connection, no battery, and no sale of surplus power — energy above the electrolyzer's rating is curtailed. This is the canonical configuration for large export-scale green hydrogen studies (IEA, IRENA, HyChain analyses).

• Onshore points compare hourly optimized solar-only, onshore-wind-only, and solar + wind designs. The result card translates normalized capacity ratios into an illustrative 100 MW electrolyzer plant; all modeled costs and outputs scale linearly.
• Offshore points model an offshore wind farm exporting power by submarine cable to an electrolyzer at the nearest shore. Cable length = great-circle shortest distance to the coastline; foundation type (fixed-bottom vs floating) follows the actual water depth at the point. Floating solar is not modeled.
• Hydrogen is delivered compressed to 200 bar into a small buffer store at the plant gate. Downstream transport (pipeline, ammonia, shipping) is out of scope.

2. Input data

3. Solar PV energy model

Annual plane-of-array irradiance is estimated from horizontal irradiance (GHI) with a transposition factor for the mounting system, plus a bifacial gain:

Energy yield applies system losses and a temperature derate computed from the site's actual air temperature (generation-weighted cell temperature ≈ annual mean ambient + 3 °C daytime bias + 25 °C irradiance rise):

Lifetime degradation (default 0.5%/yr) enters as a discounted average energy factor. Inverter clipping (DC/AC ratio) is not modeled separately — it is inside the system-loss term. The output duration curve used for electrolyzer coupling is a calibrated parametric curve g(x) = 1 − (x/x₀)ᵖ, x₀ = 0.45 whose mean equals the computed CF — i.e. generation occurs in ~45% of hours (daylight), peaking at rated power. The hybrid study instead converts each hourly GHI and temperature observation to PV output, applying the same mounting, bifacial, system-loss, temperature, and degradation assumptions.

4. Wind energy model

NASA POWER's 50 m wind speed is extrapolated to hub height with the power law v(h) = v₅₀·(h/50)^α (α = 0.14 onshore, 0.11 offshore). Speeds are assumed Weibull-distributed with shape k (2.0 onshore — the Rayleigh assumption standard when only the mean is known — 2.2 offshore) and scale c = v̄/Γ(1+1/k).

The speed distribution is pushed through a generic turbine power curve parameterized by specific power (W of rated capacity per m² of rotor): cut-in 3 m/s, cut-out 25 m/s, rated speed from P/A = ½·ρ·Cp_eff·v³ with Cp_eff = 0.42, cubic power rise between. Wake, electrical and availability losses (default 13% onshore / 12% offshore) scale the result. Sampling 400 Weibull quantiles yields the standalone screening duration curve. The hybrid study bypasses the Weibull assumption: each observed hourly 50 m speed is extrapolated to hub height and passed directly through the same turbine power curve.

5. Electrolyzer coupling — the sizing ratio

The central design decision of an islanded hydrogen plant is the sizing ratio r — generator capacity per unit of electrolyzer capacity. Oversizing (r > 1) fills the electrolyzer's duty cycle (more kg per $ of electrolyzer) but curtails generation peaks (fewer kWh per $ of generator). From the generation duration curve g(x):

Defaults are r = 1.6 for solar and 1.15 for wind; the Optimize climatology ratio button sweeps r from 0.6 to 3.0 and adopts the LCOH-minimizing value for the clicked site. Minimum-load defaults (5% PEM, 15% alkaline) are screening assumptions, not equipment guarantees. Below this threshold the standalone model produces no hydrogen. In load or dynamic mode, each productive duration-curve sample uses the load-dependent system consumption described below. Constant mode retains the original fixed-consumption baseline. Because these curves discard chronology, standalone and offshore screening do not simulate startup, standby, thermal memory or cycling wear.

Variable-power electrolyzer response

The Power-response model control in Assumptions provides three comparisons. All use the same rated system energy consumption, minimum-load limit and renewable supply. The default is the dynamic scenario; PEM and alkaline have independently editable coefficients and response times.

For productive operation at accepted fractional load q, the warm system-specific electricity consumption (SEC) is:

g is the part-load gain coefficient, a the low-load auxiliary coefficient, and c the cold-consumption penalty, all entered as percentages and converted to fractions. The thermal multiplier applies only in dynamic mode. Rated SEC remains 55 kWh/kg for PEM and 53 kWh/kg for alkaline by default. These anchors include the rectifier and balance of plant: the load curve redistributes their effect with load, so production auxiliaries are not added again. At q = 1 and reference warmth the curve returns exactly the rated SEC.

Electrical boundary: to preserve the atlas's existing capacity convention, the normalized 1 kW intake covers electrolysis and optional downstream compression together. Load q is accepted power at that shared intake divided by its rating. Production at that load is q / [SEC(q, θ) + compression SEC] kg/h per kW of intake. Thus compression is paid for from the same electricity budget, not from an additional supply. This convention approximates a package-level response; a manufacturer curve specified against stack DC power would first need conversion to the same boundary.

For reproducibility, constant mode also retains the original compressor sizing of 1 / rated SEC kg/h per kW of intake. Load and dynamic modes size the compressor from simulated peak hydrogen flow. Their LCOH comparison therefore includes this sizing refinement as well as the change in electrolyzer response.

With the default Normalized response option, warmth θ is a normalized state: 0 means ambient and 1 means reference operating temperature. It is not a prediction in degrees Celsius or a full heat balance. A start initiated at θ ≥ 0.5 uses the warm-start delay; a colder start uses the cold-start delay. The model starts cold. First-order heating and cooling use exponential updates:

Warm standby can only draw from current renewable output. With no battery or grid, a zero-power solar night cannot keep the plant warm using free energy. Startup and standby consumption are included in plant electricity use and effective kWh/kg, but never produce hydrogen. Cold and warm start counts record initiated attempts, including attempts interrupted before production.

Startup and thermal response are integrated within each weather interval while its renewable power is held constant. This accounts for a startup occupying only part of an hour, but does not reconstruct subhourly solar or wind fluctuations. No fast electrical ramp-rate limit, pressure controller, gas-purity transient or detailed thermal-control system is represented.

Research motivating these mechanisms includes Virah-Sawmy et al. (2024), whose fixed-efficiency/hourly baseline overestimated production by 5–24% relative to their variable-efficiency/five-minute cases at six Australian sites; this range is specific to their curves and baseline. Crespi et al. (2023) show why a PEM system curve must include auxiliaries, while Rauls et al. (2024) demonstrate temperature and control effects during renewable-powered PEM operation. Zhang et al. (2021) demonstrate variable-wind operation of an alkaline pilot, motivating a separate technology configuration without transferring that pilot's operating limits to all alkaline systems.

Editable equipment reference designs

The PEM and alkaline reference buttons in Assumptions select an equipment load table and an equipment heat balance. These are assumed designs, not named vendor products or fits to published experiments. Users can replace each value with data for their chosen equipment. Applying a preset replaces performance settings for that technology; costs, financing, cycling sensitivity and the other technology's values are retained. Settings are saved in the browser and included in study exports. Reset all restores the simplified default model.

All entries are editable assumed warm-system values. Interpolation between points is linear; below 5% the first value is held, but the minimum-load limit still applies. The 100% point is the rated SEC shared by all modes. Normal production BoP consumption is already included. Convert supplier data to this model's shared-intake boundary before entering it; stack-only DC data are not interchangeable with whole-system AC data. Constant mode ignores the remaining table points. Duration-curve screening uses the selected curve at reference temperature without thermal memory.

The optional thermal model uses a single equipment temperature and heat balance:

The 39.4 kWh/kg term is hydrogen's higher heating value. Retained heat represents only the portion of excess system energy entering the modeled thermal mass; the remainder dissipates outside it. Downstream compressor heat is excluded. Heating is the smaller of heater capacity and the already budgeted startup or standby draw; it is a subset of that electricity, not an additional charge. There is no dedicated production heater. Cooling removes heat toward the temperature target up to its configured thermal capacity; normal cooling auxiliaries remain inside system SEC.

PEM / alkaline reference assumptions are respectively 70 / 80 °C operating temperature, 1.5 / 3.5 kWh/K thermal capacity per MW, 120 / 180 W/K ambient loss per MW, 100 / 200 kW heating per MW, and 400 kW thermal cooling per MW for either technology. Both initially retain 80% of excess process energy and apply a 0.2%/°C cold SEC penalty. These numerical choices are assumptions. The lumped-model approach is informed by Ulleberg (2003); that reference does not validate these preset values.

Initial equipment temperature equals ambient. Valid weather temperatures drive losses; missing temperatures use the editable fallback (15 °C by default). Explicit timestamp gaps receive no power and cool using the preceding ambient value. A warm start means the equipment is at least halfway from current ambient to the reference temperature. The configured startup delay determines when production begins; there is no temperature or gas-purity interlock. Thermal updates are exact for constant heat flows; temperature-dependent SEC is evaluated at intervals of at most five minutes during production transients. This numerical resolution does not create subhourly weather data.

Diagnostics include the temperature range, startup/standby heater energy (already counted in electricity consumption), cooling heat, and approximate operating hours above target. Undersized cooling may exceed the target; temperature is not artificially clipped and no protective shutdown is assumed. Such a case needs revised equipment assumptions. Both thermal choices omit separate fluid loops, feed and gas enthalpy, pressure, and detailed controllers; neither is a validated plant model.

Hourly solar + wind optimization

For onshore points, shared plant intake capacity is normalized to 1 kW and the model jointly searches PV capacity rₛ and wind capacity rᵥ, each expressed as kW of generation per kW of electrolyzer. The hourly profiles remain in timestamp order, preserving daytime, weather and seasonal co-variation. They are never independently sorted into duration curves before being combined.

Solar, wind and hybrids use the same response simulator and financial assumptions. Renewable electricity is allocated to production (including optional compression), startup, warm standby or curtailment. Surplus above the intake rating is curtailed. Power below minimum production load can support standby in dynamic mode; otherwise it is also curtailed.

The five-year simulation is converted to a representative mean annual output using a standardized 8,760-hour year, including the elapsed duration of leap days and missing-data gaps in the denominator. Missing resource hours are conservatively treated as no supply with passive cooling; they are not interpolated or silently removed to join unrelated operating hours. This can underestimate production when coverage is incomplete, so the results report weather coverage. A missing ambient temperature observation uses 15 °C for PV conversion and the editable ambient fallback for the equipment heat balance. The simplified normalized thermal state does not use ambient degrees.

Comparison results report annual hydrogen, effective kWh/kg, starts, standby consumption, curtailment and LCOH. The numerator of LCOH separately annualizes PV, wind, electrolysis, stack replacement, water, compression and buffer-storage costs. Replacement timing follows productive operating hours, with optional start-equivalent wear as described below. Solar and wind energy shares describe available renewable generation before curtailment; consumed or curtailed hybrid electricity is not assigned back to one technology.

Weather period versus project life: 2016–2020 supplies five years of hourly operating chronology; the default 25-year project life is the financial horizon used to annualize capital and replacement costs. The model assumes the five-year weather sample is representative of operations over that horizon; it does not currently download or simulate 25 distinct weather years.

Capacity search and recommendation rule

• Pure-resource boundaries: solar-only and wind-only ratios are searched from 0.50× to 3.50× in 0.025× increments. These are valid candidates, not fixed comparison cases.
• Hybrid interior: the search starts on a 0.25× two-dimensional grid, refines the six cheapest neighborhoods at 0.05×, then refines the best neighborhood at 0.01×. Each PV and wind ratio is capped at 3.50×.
• Selection: the lowest-cost hybrid is compared with the lower-cost optimized pure boundary. A hybrid is recommended only if it is the numerical minimum and saves at least 0.5%; otherwise the best pure design is reported as the more robust recommendation.
• Changed assumptions: editing any cost or performance assumption reruns the optimizer against the already downloaded hourly weather. The hourly data are not refetched unless the selected point changes.

6. Financial method

All costs are annualized with the capital recovery factor at a single real WACC (default 7%):

• Component lifetimes differ: generation assets over their own life (30 yr PV / 25 yr wind), transmission over 30 yr, electrolyzer and compression over the 25-yr project.
• Interest during construction: CAPEX × (1 + WACC × build-years ÷ 2).
• Contingency (default 5%) applies to every direct CAPEX line.
• Stack replacement is scheduled from the stack-hour budget divided by annual productive operating hours plus optional equivalent start-wear hours. It is costed using the selected stack replacement fraction plus 10% labour, discounted to present value and annualized.
• Water is charged per kg (10 L demineralized × 1.35 raw-water factor × $/m³); demin equipment CAPEX sits inside the electrolyzer balance-of-plant line.
• Compression & storage (optional, on by default): 30→200 bar compression energy is drawn from the same renewable supply (raising kWh/kg). Load and dynamic modes size the compressor to peak H₂ flow; constant mode preserves legacy nominal sizing. The buffer store is sized in hours of average production.

Cycling degradation as a sensitivity

Cycling sensitivity: hours per start defaults to zero, so cycling does not shorten stack life unless explicitly enabled. A nonzero coefficient adds the same equivalent operating-hour charge for every cold or warm start attempt in the chronological dynamic simulation:

This changes the replacement-cost schedule only. It does not simulate increasing SEC with age or apply an extra hydrogen-yield penalty. Kuhnert et al. (2024) found that shutdown strategy affected PEM cell degradation in accelerated stress tests, but explicitly distinguish qualitative mechanisms from a quantitative plant-lifetime model. No universal hours-per-start coefficient can be inferred from that experiment. The sensitivity is therefore a user-selected scenario, not an empirical lifetime forecast; duration-curve screening cannot compute it because it contains no start chronology.

7. Offshore transmission

For offshore points the export system is sized to the electrolyzer intake (excess wind output is curtailed at the platform, so the cable never needs to carry the full oversized farm rating). The model prices both options at the site's cable distance and picks the cheaper:

The crossover lands near 70–90 km, matching industry practice. Cable losses reduce delivered energy. Foundations: fixed-bottom to 60 m depth (ETOPO1), floating beyond — floating adds platform/mooring CAPEX, extra installation cost and higher O&M. The cable route is drawn as a straight line to the nearest shore point; real routing, seabed conditions and landfall constraints are site-specific.

8. Every default, with its source

The tables below are generated from the same file the model reads (assumptions.js) — the documentation cannot drift from the calculation. All values are editable in the app's Assumptions tab. Costs in 2024 US$; rates are real. The new electrolyzer response coefficients and timing defaults are explicitly labeled illustrative scenario assumptions; supporting papers establish mechanisms rather than a universal default parameter set.

9. Known limitations

• The onshore hybrid study covers five historical weather years, not the full 20-year climatology; longer samples and independent ERA5 validation would improve weather-risk analysis.
• No battery or chronological hydrogen-storage dispatch. Buffer storage is costed from hours of average production but does not impose a constant delivery schedule.
• Electrolyzer load curves, startup timing and thermal response are generic scenario assumptions, not calibrated equipment predictions. The normalized proxy and optional lumped heat balance omit pressure dynamics, detailed fluid-loop heat balances, gas purity, modular stack dispatch and fast electrical ramp limits. Optional cycle wear changes replacement cost but is not a validated plant-lifetime model.
• Hourly renewable inputs cannot resolve fast cloud or wind fluctuations. Integrating startup within an hour does not add weather information. Missing resource hours are treated as zero supply, and results should be interpreted alongside the reported coverage.
• The hybrid assumes PV, wind and electrolysis are co-located behind one unconstrained electrical bus; internal collection networks, generation-to-electrolyzer transmission losses and spatial separation between the best solar and wind sites are not modeled.
• Hourly PV output uses GHI with a single transposition multiplier rather than a solar-position and plane-of-array model; hourly wind uses a generic turbine curve rather than a site-selected machine.
• The standalone/offshore model uses Weibull-from-mean wind estimation. Neither model includes terrain speed-up or site-specific roughness.
• No land-use, protected-area, water-availability or grid-proximity screening — a cheap cell in the Sahara still needs a port and water; the water cost is included, its availability is not.
• Single global WACC unless you change it — in reality financing costs vary by country more than technology costs do.
• Offshore: straight-line cable route; electrolyzer assumed onshore (in-turbine or platform electrolysis with H₂ pipelines to shore is an emerging alternative not modeled).
• No hydrogen delivery beyond the plant gate, no conversion to ammonia/derivatives.

10. References

• IRENA, Renewable Power Generation Costs in 2023 (2024) — installed cost, O&M and LCOE benchmarks for solar PV, onshore and offshore wind.
• NREL, Annual Technology Baseline 2024 — bottom-up CAPEX breakdowns, O&M, lifetimes and financial assumptions.
• IEA, Global Hydrogen Review 2024 — electrolyzer system costs, efficiencies and deployment status.
• US DOE Hydrogen & Fuel Cell Technologies Office — H2A production records and technical targets (electrolyzer CAPEX shares, stack life, compression energy and costs).
• Danish Energy Agency, Technology Data catalogues — generation and renewable-fuel plant cost/performance data.
• NREL offshore wind cost studies and ORE Catapult guides — offshore CAPEX splits, export-system (HVAC/HVDC) costs and the AC/DC distance crossover.
• Lazard, LCOE+ (June 2024) — cross-checks on LCOE and hydrogen cost ranges.
• NASA POWER, API documentation; Jordan & Kurtz (NREL), Photovoltaic Degradation Rates; IEA Wind Task 37 15 MW reference turbine.
• Virah-Sawmy, D., Beck, F. J. & Sturmberg, B. (2024), Ignore variability, overestimate hydrogen production – Quantifying the effects of electrolyzer efficiency curves on hydrogen production from renewable energy sources, International Journal of Hydrogen Energy 72, 49–59. Supports load-dependent efficiency and timestep sensitivity; its reported production differences are not generic penalties.
• Crespi, E., Guandalini, G., Mastropasqua, L., Campanari, S. & Brouwer, J. (2023), Experimental and theoretical evaluation of a 60 kW PEM electrolysis system for flexible dynamic operation, Energy Conversion and Management 277, 116622. System boundary, auxiliaries and part-load behavior; not the source of this atlas's curve coefficients.
• Rauls, E. et al. (2024), System dynamics of polymer electrolyte membrane water electrolyzers and impact of renewable energy sources on systems design, International Journal of Hydrogen Energy 65, 83–94. Experimental thermal and pressure responses on a 100 kW test station.
• Zhang, C. et al. (2021), Wind-powered 250 kW electrolyzer for dynamic hydrogen production: A pilot study, International Journal of Hydrogen Energy 46, 34550–34564. Alkaline operation with variable wind inputs; demonstrated limits belong to the tested pilot.
• Kuhnert, E. et al. (2024), Impact of intermittent operation on photovoltaic-PEM electrolyzer systems: A degradation study based on accelerated stress testing, International Journal of Hydrogen Energy 55, 683–695. Motivation for a cycling sensitivity, not calibration of a transferable degradation-per-start coefficient.

← Back to the map