The French offshore wind sector is undergoing rapid expansion, with a projected cumulative capacity of 3.6 GW expected by the end of 2027, driven by the commissioning of seven major offshore projects. This growth occurs in a context that is both innovative and promising, yet increasingly competitive.

Long-term durability and reliability of offshore wind structures critically depend on their ability to withstand evolving marine conditions over operational lifespans that often exceed two decades. Climate change is modifying key metocean variables, including significant wave height, total sea level, and wind speed, as well as the frequency and intensity of extreme events (). These evolving conditions introduce substantial uncertainty in design parameters, particularly in regions exposed to high wave energy and strong winds, which remain the most favorable for energy production (). Neglecting these changes may result in under-designed structures, reduced operational lifespans, increased maintenance costs, or even catastrophic failure ().

Metocean studies are essential for understanding and forecasting the environmental conditions that offshore structures will face throughout their service life. In particular, extremes in significant wave height, total sea level, and wind speed are critical design parameters that directly influence structural performance, safety, and operational efficiency (). Significant wave height, representing the average height of the highest third of waves, is a key factor in determining wave loading on structures and plays a central role in fatigue analysis, foundation stability, and the overall design of wind turbine components ().

Traditionally, offshore wind design assumes stationarity, meaning that the statistical properties of environmental conditions (mean, variance, extreme values…) are constant over time. However, the non-stationary nature of climate change challenges this approach. Sea level rise, shifting storm tracks, and intensifying wave conditions can significantly affect wave loading, fatigue life, and structural integrity . Capturing non-stationary behavior is essential for robust design and accurate estimation of future extreme conditions . Seasonality is also crucial, as more intense winter extremes and calmer summer conditions may affect additional fatigue (). Correctly characterizing these contrasts is critical for fatigue analysis, foundation stability, and wave loading calculations (). Although non-stationary GEV models are widely studied in the statistical literature (), their application to CMIP6 climate projections for French offshore sites is limited. And projecting return levels up to 2100 raises additional challenges of data extrapolation and model uncertainty, as estimating long period return values in a changing climate is known to be associated with large uncertainties related to model variability and methodological choices (). So, existing methods relying solely on historical stationary data may underestimate future extremes, leading to inadequate design loads.

Neglecting these evolving trends may lead to under-designed structures, reduced operational lifespans, and increased maintenance costs, ultimately compromising the economic viability and safety of offshore wind projects. It is therefore imperative to assess the impact of climate change on sea state conditions to inform the design criteria of offshore wind farms and other coastal infrastructure. This research aims to address this challenge by suggesting a new methodology to better investigate projected changes in significant wave height. The final objective is to provide engineers, designers, and stakeholders with updated tools that account for non-stationary environmental conditions, ensuring that offshore wind farms remain resilient and cost-effective in a changing climate.

This study develops a statistical framework using non-stationary GEV models applied to monthly maxima to better quantify projected changes in significant wave height. Its main contribution lies in applying these models to CMIP6 projections along the French coasts to assess future extreme sea states, while also deriving an equivalent design level that accounts for non-stationarity over the operational lifespan of offshore structures, following the approach recently formalized by . This framework provides engineers, designers, and stakeholders with updated tools to ensure resilience, safety, and economic viability of offshore wind farms under evolving climate conditions. In this paper, the statistical methodology is thoroughly presented in Sect. 3, focusing on the application of non-stationary GEV models to monthly maxima. Subsequently, Sect. 4 highlights the results, emphasizing projected changes in return levels and extreme sea states extending up to 2100. Finally, Sect. 5 discusses the implications of these findings for offshore wind design and adaptation strategies, while critically examining the uncertainties associated with long-term extrapolation and model assumptions.

We now present the results obtained from fitting the non-stationary models to the wave data. For clarity, we report detailed findings only for a representative location in the Eastern English Channel; results for all other sites are provided in the Supplement.

Figure  displays the 99th quantile, calculated using Eq. (), and estimated for each GCM at the East English Channel location. As stated in Sect. , this quantity is dependent on the month and on the year, as a function of both the position and scale parameters. The years are represented on the x-axis, while the y-axis contains the months, from January (bottom of each subplot), to December. It is important to notice in this figure that the time period is discontinuous between the historical and future period.

The figure clearly illustrates the pronounced seasonality of significant wave height (Hs), with consistently higher values during winter (yellow) compared to summer (blue) across all models and time periods, suggesting a shift in seasonal storm patterns. A seasonal contrast in Hs emerges between the historical period and scenario SSP1-2.6, and becomes even more pronounced under SSP5-8.5. Historical Hs values generally range from 3 to 7 m, with gradual transitions represented by thick bands in Fig. 4. Under SSP1-2.6, the range remains similar (3 to 7 m), but seasonal variability increases, as indicated by narrower bands. In contrast, SSP5-8.5 exhibits a wider range (2 to 8 m) and sharper seasonal fluctuations, with Hs shifting from about 2.5 m in August to 5 m in September, for example. These results suggest a trend towards more intense seasonal contrasts, characterized by higher winter values (up to 8 m) and lower summer values (around 2 m), particularly evident when comparing the historical period to SSP5-8.5.

Furthermore, the figure reveals a time-evolution of the seasonality: for most models, the summer period tends to become longer and to peak later in the year, though exceptions exist (AWI, KIOST, and MRI models). Finally, the models successfully reduce to the stationary case when the data shows no evidence of shifting seasonality (e.g., KIOST model under SSP1-2.6).

An additional remark can be made regarding the results of MRI-ESM2-0 under scenario SSP1-2.6, for which the seasonality is very pronounced in the projection period, with a very large shift in the seasonal cycle (see Fig. ). This behaviour is also observed at the other locations (see the Supplement). As this behaviour is unexpected and cannot be explained by any physical mechanism, this model should perhaps be disregarded when drawing a precise assessment of the evolution of extreme sea states under climate change. However, since the focus of this paper is on the methodology, we decided to retain this model in the remainder of the analysis.

Figure  illustrates the shift in seasonality discussed earlier by showing, for each model, the monthly extreme quantiles estimated with the non-stationary approach for two reference years: 2010, corresponding to the end of the historical period, and 2100, representing the end of the future period. The most notable feature highlighted by this figure is the seasonal shift observed in most models when comparing the grey curve (2010) with the colored curves (2100 under two scenarios). The latter exhibit lower values during summer and higher values in winter. Another striking aspect is the asymmetry between the beginning and end of the year in future scenarios, with January values exceeding December by approximately 0.2 to 0.5 m, depending on the model. Furthermore, the minimum Hs values tend to occur later in the year, and the adjacent slopes become steeper, indicating a more pronounced and extended summer season compared to the historical period. Overall, all models suggest a shift of the seasonal cycle toward the end of the year, implying a potential redefinition of conventional seasons. For instance, the lowest Hs values now occur in July rather than June. Finally, the comparison between SSP1-2.6 and SSP5-8.5 reveals that the latter amplifies these effects. While all models generally agree on this trend, the KIOST model shows the smallest changes, whereas the MPI model exhibits the largest.

Because of the shifts identified both in the summer and winter values, it may be difficult to assess if the annual return levels also evolve with time, because the two effects can compensate each other. This can be investigated in Fig. . To compute the annual return level from the fitted model, we first compute the distribution of the annual maxima, using Eq. () of Sect. , by taking the product of the CDF of monthly maxima. This leads to the CDF of the annual maxima, that is then inverted numerically to obtain the return levels, which still depends on the year. Uncertainties are estimated using a Monte-Carlo approach, by simulating from the fitted model.

We see that the shift of the summer and intensification of the winter may lead to an increase of the annual return level for AWI, MRI and also MPI but for SSP5-8.5 only. While some models as ACCESS-CM2, EC-EARTH, IPSL and KIOST can also indicate that there are no long-term evolution of the return levels. And finally, the CMCC model shows a contrasted result with an increase of the return level under SSP1-2.6 and a decrease under SPP5-8.5.

We can also notice a more complex pattern in the MPI model during the historical period, which would not have been captured by a classical linear approach. Looking back at Fig. , this behavior results from a shift of the summer season towards the beginning of the year, followed by an intensified winter that fails to compensate, producing an “inverted bell curve”.

In order to compare the estimated values to the ones obtained using a classical stationary model, we estimated the equivalent design condition using Eq. (). In a stationary climate, this definition would coincide with the 100-year return level. These quantities are compared on Fig. , for the different models and scenarios. We included the 100-year return level estimated from annual maxima (points), monthly maxima (triangles) and the proposed non-stationary approach (squares).

The main insight from this figure is that using monthly maxima instead of annual maxima significantly reduces the uncertainty in estimating 100-year return levels. This improvement is largely due to the model being fitted to twelve times more data, despite its increased complexity. The figure also shows strong consistency across methods, with estimates generally falling within each other's confidence intervals. In contrast, the annual maxima approach may yield spurious results because of the limited data available for fitting the GEV model in this case.

For stationary cases as identified before (e.g. KIOST model under scenario SSP1-2.6), all three approaches produce similar estimates, with the fully non-stationary model providing the narrowest confidence intervals. When averaging return levels across all models, we obtain approximately 6.15 m for the historical period, 6.75 m for SSP1-2.6, and 7 m for SSP5-8.5, indicating that future scenarios predict higher return levels.

Lastly, Fig.  contains the estimation of the equivalent design condition for each seafront representative point detailed previously (Table ), and using the methods used throughout the paper. All trends shown are progressive, first between the historical period and SSP1-2.6, then with SSP5-8.5, even though the return periods are very similar for the future scenarios. Only the East Channel location shows a slight decrease, the other representative points tending toward an increase. West Channel location has a slight increase, North Atlantic location has a clear increase of approximately 2.5 m, South Atlantic location has an increase of approximately 1 m, West Mediterranean location has a significant increase of 5 m for the historical scenario and 10 m for future scenarios, and finally East Mediterranean location has an increase of 1 m for the historical scenario and SSP5-8.5 and 2 m for SSP1-2.6. Globally, we can derive the same comments as the previous figure for points located in the English Chanel and in the Atlantic. The non-stationary monthly method reduces uncertainties and has values close to those of the monthly maxima method. However, for representative points in the Mediterranean Sea, the annual maxima method gives the lowest margin of uncertainty. It should be noted that West Mediterranean location has very high Hs values compared to all other representative points, reaching up to 20 m and uncertainties of up to 30 m. For the locations in the Mediterranean sea, the results are more mitigated compared to the English Channel and the Atlantic, which can be explained by a less important seasonality (see Figs. S9 and S10 in the Supplement) and to a less-convincing representation of the sea-states from the GCMs in this area. In fact, the coastal processes are very important and not well represented in the very large-scale climate models used to force the numerical wave model. This has a particular impact on the Mediterranean Sea, which is surrounded by coastlines, compared to the Atlantic Ocean or the English Channel.

To conclude, the lifetime design level represents the wave height to not exceed within the structure lifetime of the wind farms. In Fig. the lifetime design level is represented for the English Channel. For all models the lifetime design level is smaller for the historical scenario from 1985–2014 (7.4 m) than for future scenarios from 2081 to 2100, the SSP1-2.6 (8 m) and SSP5-8.5 (8.2 m) are close to each other. The largest uncertainties are for SSP5-8.5. The models with the highest uncertainties are CMCC and MPI for the historical period and SSP5-8.5 and it is KIOST for SSP1-2.6. The investigation of the lifetime design level is interesting through all maritime seafronts (see Supplement – Figs. S13 to S16). The Atlantic coast presents similarities with the English Channel even if values are higher with 11 to 13 m for the historical period and 13 to 15 m for both SSPs. The South part of the Atlantic has the lowest values compared to the North part. And for the Mediterranean Sea SSP1-2.6 seems higher than SSP5-8.5, showing really high uncertainties even if compared to all other seafronts. For the eastern part of the Mediterranean Sea, values are low around 10 m for the historical period and 12 m for the two SSPs, whereas for the West part values are high, around 20 m for all scenarios. The Mediterranean Sea highlights great uncertainties from 4 to 14 m so around 10 m difference between all models for SSP scenarios while uncertainties are lower in the Atlantic and the English Channel with a maximum of 3 m difference between all models.

The three French maritime seafronts – Atlantic, English Channel, and Mediterranean – display fundamentally different projection behaviors and uncertainty envelopes in climate-driven wave simulations, with significant implications for offshore wind development. Ensemble analyses and regional studies show that the Atlantic and Channel coasts generally produce more coherent signals across models, notably clearer changes in storm-track activity and a pronounced intensification of winter extremes, whereas the Mediterranean sea emerges as both a climate hotspot and as a basin with much larger inter-model spread and several physically implausible outliers in some model-wave ensembles (e.g., extremely large simulated Hs in a few runs) which are likely linked to coarse GCM resolution, poor representation of local cyclogenesis, and unresolved coastal/fetch effects (). Several authors therefore caution that Mediterranean projections must be interpreted with care: ensemble medians often differ from individual model extremes, and values that appear unrealistically high typically arise from biased wind forcing or from insufficiently resolved dynamical drivers rather than from robust physical signals (). By contrast, the Atlantic basin frequently shows a distinct “deepening” of seasonality in extremes, a stronger winter peak and starker summer contrast in Hs, a signal that is reproduced more consistently across multiple modelling studies and wave projections (; ). Finally, CMIP6 (and earlier CMIP5) model families exhibit substantial inter-model differences in the representation of the large-scale drivers of wave climate (storm tracks, surface wind fields and seasonal variability), with specific models (e.g., KIOST-ESM, some CMCC configurations, MPI-ESM variants) producing divergent wind-forcing fingerprints and therefore divergent wave projections; this heterogeneity has been documented in model evaluation and ranking studies and argues for multi-model, bias-corrected and high-resolution downscaling strategies when deriving local extreme Hs estimates (). In summary, while Atlantic and Channel projections provide relatively coherent seasonal signals – particularly a clearer intensification of winter extremes – Mediterranean projections remain more uncertain and prone to isolated, physically questionable extremes. Risk assessments in the Mediterranean sea therefore demand rigorous model selection, downscaling, and plausibility checks rather than blind reliance on raw GCM-driven wave outputs ().

The seasonal variability of extreme significant wave heights shows marked contrasts between France's three maritime seafronts, reflecting both atmospheric circulation patterns and basin morphology. Along the Atlantic coast and in the English Channel, studies by and highlight a clear winter dominance of extreme events, associated with intensified westerly storms and North Atlantic depressions, whereas summer remains generally calmer, confirming a strongly seasonal regime. These works suggest that winter maxima are projected to intensify consistent with our results suggesting the increase of severe winters and really low values of Hs in summer. In fact, anticipate an increase of annual maxima Hs in the three French seafronts for mid-century and end of century with RCP4.5 and RCP8.5. Overall, all studies agree that the cold season remains dominant for extreme Hs, even though future projections point to a more complex seasonal cycle, with longer and later summers. , using ERA5 from 2001 to 2020, found that Hs 95 percentile values are increasing in the North Atlantic coast and in the South of the English channel. As discerned a shift towards higher extremes in a future wave climate similar to our results in the North Atlantic with RCP4.5 and RCP8.5 comparing 1970–1999 to 2071–2100. An intensification of the 100 return level is found in the Atlantic and the English Channel by using the POT and GPD using numerical wave hindcast BOBWA database from 1958 to 2002 in accordance with our results. Within this context, our results showing a tendency towards more intense winters and calmer yet delayed and prolonged summers, together with an increase in the return period of 100-year Hs are consistent with these previous findings, suggesting a future regime characterized by stronger seasonal contrasts and a temporal redistribution of calm conditions towards early autumn.

The adoption of non-stationary extreme-value models has emerged as a critical advance in assessing changes in wave extremes, offering superior performance compared to stationary assumptions under evolving climate conditions. Traditional stationary models presume that the statistical characteristics of wave climate remain constant over time, an assumption increasingly invalid in the presence of long-term trends, changing storm regimes or evolving wind and fetch patterns as there are important in French seafronts. showed that including time-varying location and scale parameters in extreme-value modelling substantially improves the estimation of return levels (e.g., centennial Hs) under non-stationary forcing and reduces bias in projections. For RCP4.5, reported an uncertainty of 13.15 m for the 100-year return level, which is substantially larger than the uncertainty obtained with our method (maximum of 10 m). This reduction in uncertainty is likely attributable to the larger amount of data used in our analysis, as the increased sample size across all models leads to more robust extreme value estimates. highlighted that non-stationary approaches can account for local, seasonal and directional variability in wave extremes across basins (e.g., North Atlantic, Mediterranean) where stationary assumptions fail to reproduce observed shifts in intensity, timing and frequency of events. More recently, emphasized that non-stationary extreme-value analyses are indispensable for climate adaptation and coastal engineering, since they quantify how the probability of rare, extreme wave events evolves under future scenarios rather than assuming historical stability. Complementing these findings, earlier studies () illustrated how modelling non-stationarity in seasonal parameters yields more realistic return value estimates for marine extremes by capturing seasonal shifts in storminess and fetch, thereby supporting the use of non-stationary parameterization for extreme Hs. Altogether, these works converge on the conclusion that non-stationary models not only offer a more realistic statistical representation of changing wave climates but are also essential for deriving credible return periods and extreme-value projections in a non-equilibrium climate system. Our study on wave extremes therefore provides a new, effective method for monthly maxima with reduced bias and taking climate change into account. Using monthly maxima instead of annual maxima provides a finer temporal resolution that better captures intra-annual variability in wave extremes and their seasonal drivers. This approach allows the detection of shifts in the timing and intensity of extreme events, an information that is lost when only annual maxima are considered. Moreover, monthly maxima improve the robustness of non-stationary analyses, as they provide a larger sample size and make it possible to explicitly model seasonal covariates (e.g., storm seasonality, climate indices), leading to more accurate and physically meaningful estimates of extreme return levels.

This study introduces a non-parametric approach based on monthly maxima to estimate return levels for monthly and annual periods from 8 GCMs, explicitly accounting for climate change effects. A key contribution lies in proposing a new method for calculating design levels – equivalent to return periods – under non-stationary conditions driven by a changing climate. The consistency of our results with existing literature reinforces the credibility and the relevance of the approach.

By leveraging monthly maxima, we substantially reduce uncertainty in extreme-value estimation. The parsimonious, data-driven model captures seasonality and its potential shifts, as well as changes in extremes themselves, making it particularly suited for assessing rare events critical to offshore wind farm design.

Our findings reveal coherent signals of stronger seasonal contrasts under future climates, with intensified winters and calmer yet prolonged summers, especially in the Atlantic and English Channel. Despite significant inter-model uncertainty – most pronounced in the Mediterranean sea – the ensemble generally indicates higher extreme sea states under both SSP1-2.6 and SSP5-8.5 scenarios.

This work opens several avenues for future research: addressing biases in extreme-value estimation to improve historical consistency; refining wave models along the French coastline for better representation of local dynamics; developing parametric approaches to seasonality and its evolution; and exploring joint distributions of significant wave height (Hs) and peak period (Tp) under future climate scenarios to enhance the robustness of offshore wind design.