Scientists put a giant floating wind platform on a “diet” and taught an AI to watch the grocery bill
Floating offshore wind grew fast — from a single full-scale prototype in 2009 to 88 MW of installed capacity by 2023. Then it stalled. Rising costs and the gap between promising technology and commercial-scale economics put the brakes on.
At the center of that gap sits the floating platform: a massive semi-submersible steel structure designed to hold a 15 MW turbine stable in deep water. It’s heavy, expensive, and exposed to volatile raw material markets. A new optimization framework may offer a way to attack both problems at once — trimming steel mass through engineering while using machine learning to bring discipline to the cost uncertainty surrounding it.
Floating wind’s cost problem
The technology has proven it can work. The harder question is whether it can work cheaply enough. Floating offshore wind needs to reduce its levelized cost of energy to compete in renewable energy auctions — and costs have moved in the wrong direction, rising alongside broader macroeconomic pressures. That squeeze has slowed the path from demonstration projects to commercial gigawatt-scale deployment.
The best-performing design reduced steel mass from 3,916 tonnes to 3,273 tonnes — roughly a 16% reduction — while satisfying every dynamic performance constraint.
Among the many cost drivers, the floating substructure stands out. Research from the Carbon Trust identified the floating platform as the highest-priority area for capital cost reduction — potentially delivering a 16% CAPEX reduction as projects move from prototype to commercial scale. Steel price volatility makes that challenge harder still. When raw material prices swing sharply, long-term project planning becomes a financial guessing game.
Engineering the lightest possible platform
The study took the IEA 15 MW reference turbine — a standard benchmark in offshore wind research — and mounted it on the UMaine VolturnUS-S semi-submersible platform, giving researchers a well-documented starting point with known performance characteristics.
A genetic algorithm then went to work, iteratively adjusting six geometric parameters: the diameters of the central and outer columns, the height, width, and thickness of the pontoons, and the overall platform diameter. Each candidate design was evaluated using RAFT, an open-source frequency-domain code, against performance constraints — including platform pitch, nacelle acceleration, tower-base bending moment, and static stress at the central column-pontoon joint.
The algorithm ran 10 separate times to account for its inherent randomness, and results were consistent. The best-performing design reduced steel mass from 3,916 tonnes to 3,273 tonnes — roughly a 16% reduction — while satisfying every dynamic performance constraint. The optimized geometry featured slightly larger outer columns and pontoon dimensions but a smaller overall platform diameter.
Teaching a model to forecast steel prices
Steel is the dominant cost driver for the platform, and its price has been anything but stable. The COVID-19 pandemic triggered sharp swings in commodity markets, and steel was no exception. Predicting where prices will go — over the multi-year window between feasibility study and final investment decision — is genuinely difficult.
The team built a hybrid forecasting tool to tackle this. It combines XGBoost, a machine learning method well-suited to capturing short-term stochastic fluctuations, with SARIMAX, a statistical model better at tracking long-term seasonal trends. Neither approach alone is sufficient — the hybrid draws on the strengths of both.
To improve accuracy, the team filtered a range of external economic indicators and ranked them as supporting inputs. Rebar prices, copper, coal, and London Metal Exchange indices all passed statistical screening and were incorporated. These variables correlate with steel prices, giving the model additional signal beyond steel’s own history.
In hindcast testing across 2022–2024 — a particularly turbulent period for commodity markets — the hybrid model outperformed each individual model. It produced the lowest mean absolute error, with residuals clustered close to zero. That’s the clearest sign a forecast is neither systematically over- nor underestimating.
What the numbers reveal about cost and viability
The economic results make the stakes clear. Steel price uncertainty alone drives a CAPEX swing of roughly €150–200 million at the 1 GW farm level — a 40–50% variation across forecast scenarios. That’s a wide spread for any investment committee to absorb.
The optimized platform design delivers consistent savings across all price scenarios: approximately 10–15% in platform cost compared to the reference design. That figure holds whether steel is cheap or expensive, making structural optimization a reliable lever regardless of market conditions. On levelized cost of energy, the picture is more moderate. LCoE varies by €3–5 per MWh across forecast scenarios, with structural optimization contributing roughly €1 per MWh of reduction in high-price environments. As the authors note, economic forces emerge as the primary driver of uncertainty, with design optimization offering meaningful but partial mitigation.
What comes next for floating wind economics
The framework isn’t locked to a single design. The authors built it to be applicable to any semi-submersible platform, meaning the methodology could be deployed across different reference designs or site-specific configurations as the industry matures.
Several refinements could sharpen the results. More detailed dynamic stress analysis would add confidence to the optimized geometry. Incorporating site-specific environmental conditions would make design constraints more realistic. On the forecasting side, longer training datasets and better-calibrated SARIMAX models could improve short-term prediction confidence, where the model currently shows some systematic bias.
Combining structural design optimization with commodity price forecasting in a single framework isn’t yet standard practice in floating wind feasibility studies. If developers start treating these two levers as inseparable parts of the same analysis, it could improve how the industry prices risk, structures procurement, and makes the case to investors that commercial-scale floating wind is financially viable.
More information is available here: White, C., Benifla, V., Cândido, J., and Gato, L. M. C.: Economic and design optimization of a 15 MW floating offshore wind platform using time series forecasting, Wind Energ. Sci., 11, 3703–3717, https://doi.org/10.5194/wes-11-3703-2026, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
