Offshore Wind Resource Estimation via HF-Radar ANN Inversion -- Vilano (Spain) Case Study

Context and Motivation

Offshore wind resource assessment is critical for wind energy development, but direct long-term measurements are often sparse. High-frequency (HF) coastal radars offer a potential alternative by indirectly sensing sea-surface winds through their backscatter signals. In the Vilano case study (Galician coast, NW Spain), we leverage an artificial neural network (ANN) that inverts HF radar observations into near-surface wind estimates. This approach (developed under the HF-EOLUS project, Zenodo DOI: 10.5281/zenodo.17464519) enables the retrieval of 10 m wind fields over the coastal ocean by combining HF radar data with limited satellite and buoy observations. The motivation of this case study is to evaluate how well HF-radar--derived winds can characterize the offshore wind resource at a site of interest (the Vilano-Sisargas buoy location) and to demonstrate a workflow for deriving wind resource metrics from remote sensing data.

Data Inputs

The analysis uses two primary datasets (packaged as STAC-compliant GeoParquet files for reproducibility):

• HF-Radar ANN Wind Inversion Dataset -- Hourly 10 m wind vector estimates (speed and direction) derived from HF radar measurements via ANN inference. These data cover a grid of 56 offshore points (mesh nodes) around the study site and constitute the core input for wind resource computations (Zenodo DOI: 10.5281/zenodo.17464583).

• Vilano Offshore Buoy Measurements -- In-situ wind observations from the Puertos del Estado Vilano buoy (Sisargas buoy), which is a moored oceanographic buoy near Cape Vilán, Spain. The buoy provides hourly wind data (at 3 m anemometer height, extrapolated to 10 m) used as an independent reference for validation (Zenodo DOI: 10.5281/zenodo.17098037).

Each dataset above is publicly available and includes spatiotemporal metadata in an STAC catalog. The ANN dataset represents a hindcast of wind conditions obtained by applying the trained ANN model to historical HF radar records, while the buoy dataset represents the ground-truth wind climate at the site. By comparing these, we ensure the ANN-derived winds are physically consistent and quantify any biases.

The derived wind resource indicators (power density, capacity factor, Weibull diagnostics, and bootstrap summaries) together with their STAC packaging are openly released as HF-EOLUS. Task 5. Wind Resource Estimation Results (Zenodo DOI: 10.5281/zenodo.17594220). Please cite this record whenever you reuse the Vilano case study outputs.

Methodology

We applied the HF-EOLUS Wind Resource Toolkit (Zenodo DOI: 10.5281/zenodo.17591545) to compute standard wind resource metrics from the ANN wind dataset, with optional validation against the buoy record. The ANN outputs (10 m height wind estimates) were first adjusted to a representative wind turbine hub height (110 m) using a neutral logarithmic wind profile (open-sea roughness length 0.2 mm). At each analysis grid point, the toolkit evaluated long-term wind statistics, including mean wind speed, percentile speeds (e.g., P90), and the wind speed frequency distribution. A Weibull distribution was fitted to the wind speed data, with appropriate treatment of left-censoring beyond the radar's reliable speed range (~5.7--17.8 m/s) to account for the ANN's classification limits. From the wind speed distribution, we computed the wind power density (the mean available wind energy per unit area, W/m²) and estimated a capacity factor by applying a standard 6 MW offshore turbine power curve to the wind data. All calculations include quality-control filters: for example, data segments with low radar coverage or any bias flags were noted (though included in aggregate metrics), and "out-of-range" wind estimates (beyond the ANN's training range) were handled via a capped distribution approach. We also conducted a block-bootstrap analysis (500 resamples for the whole dataset) to quantify statistical uncertainty in key metrics (e.g., confidence intervals for mean wind speed and power density). Finally, for validation, the toolkit synchronised the ANN predictions at the buoy's location with the buoy's observed time series, enabling a direct hour-by-hour comparison of wind speed and direction.

Note: The methodology strictly avoids site-specific tuning beyond the ANN model itself -- all parameters (e.g., vertical extrapolation, censoring thresholds) follow standard or documented values to ensure the results are generalisable. No explicit wave or stability corrections were applied (a neutral atmosphere was assumed), and the capacity factor estimation is theoretical (no wind farm losses or wake effects were included), consistent with a resource assessment scenario.

Results

Wind Resource Characteristics: The ANN-based wind climatology indicates a strong offshore wind resource in the Vilano region. Table 1 summarizes the key wind resource metrics at 110 m height, aggregated over the 56 analysis nodes. The median hub-height wind speed is about 9.4 m/s, and even the lower-ranked sites exhibit mean speeds on the order of 9 m/s. The upper end (90th percentile among sites) of mean wind speeds is about 10.0 m/s, reflecting consistently high winds across the area. These winds translate to a high available energy density: the median wind power density is roughly 785 W/m², with the windiest locations reaching ~1000 W/m². Such conditions would correspond to a substantial energy yield -- for a generic 6 MW offshore turbine, the median capacity factor is estimated around 0.58 (i.e. 58% of maximum output on average), with the best sites approaching 0.69. These figures underscore the excellent wind potential of this coastal area.

Table 1: Long-term wind resource metrics derived from the HF-radar ANN dataset for the Vilano study area. Median values (typical site) and P90 values (top-performing sites) are shown. Wind power density is the mean available power per unit area of wind, and capacity factor is the fraction of a 6 MW turbine's output that would be realized on average under these winds.

The ANN-inferred wind regime is not only strong on average but also shows relatively robust high-wind occurrence (P90 levels are close to the medians, indicating a narrow spread of consistently high winds). This consistency is partially a result of the censoring -- extremely low or high winds fall outside the radar's optimal sensitivity, so the effective distribution is focused in the moderate-to-high range. Despite this, the bootstrap uncertainty analysis suggests that for sites meeting basic data-quality thresholds, the uncertainty in mean wind or power estimates is on the order of ±10--15% (95% confidence), which is acceptable for preliminary resource evaluation.

Validation against Buoy Measurements: We compared the ANN-derived wind time series with the buoy's observed winds at the buoy location (both adjusted to 110 m height). Over the overlapping data period (11,412 hourly pairs), the ANN consistently predicts higher wind speeds than those recorded by the buoy. For example, the ANN-predicted mean wind speed at the buoy site is about 11.3 m/s, whereas the buoy's measured mean is 7.8 m/s over the same hours. Similarly, the ANN's 90th percentile wind speed is 15.4 m/s, compared to the buoy's 13.5 m/s. Despite this positive bias in wind speed, the wind power density estimates from ANN and buoy are more aligned -- 627 W/m² vs 607 W/m² at 110 m, respectively, for the matched period. Table 2 presents these comparisons, along with the buoy's long-term climatology for context. Notably, the buoy's full-record mean wind speed is 8.9 m/s (higher than the 7.8 m/s during matched hours), and its overall power density is ~850 W/m², indicating that some high-wind events recorded by the buoy were likely missing in the ANN dataset. Even so, the ANN captures the general wind power potential well, slightly overestimating speed but yielding power density close to the buoy's observations.

Table 2: Wind climate comparison at the Vilano buoy location. "Paired hours" refers to the time span where both ANN and buoy data are available simultaneously. For reference, the buoy's full climatology (all available data over the deployment period) is summarized on the right. The ANN shows a high bias in wind speed relative to the buoy, but the computed power densities are within ~3% during matched periods. The buoy's full record includes some higher-wind periods that were not captured in the ANN input, hence its higher long-term power density.

This validation highlights some important considerations. The ANN model tends to overestimate wind speeds at this specific location. Nevertheless, the fact that power density -- a function of the cube of wind speed -- is so close between ANN and buoy suggests that the overall energy content of the wind is being captured reasonably well by the ANN inversion. The buoy's higher long-term power density (850 W/m²) compared to the paired-sample value (607 W/m²) is mainly due to periods of very high winds that the radar missed; excluding those, the ANN-derived dataset provides a representative picture of the wind resource.

Conclusion

This case study demonstrates a novel use of HF radar data for offshore wind resource estimation. By applying an ANN inversion and a rigorous analysis toolkit, we obtained detailed wind climatology metrics for the Vilano offshore site without relying on dedicated meteorological masts or Lidar. The results show that ANN-inferred HF radar winds can reliably reproduce key wind resource indicators, such as mean wind speed and energy density, in a coastal environment. The median wind speeds (~9--10 m/s at 110 m) and high capacity factors (~60%) highlight the promising wind energy potential of the Galician coastal waters. At the same time, validation against buoy data reveals limitations. There is a systematic high bias in ANN wind-speed estimates and considerable scatter in wind direction, underscoring the importance of thorough calibration and uncertainty analysis. Overall, the Vilano case study provides proof of concept that HF radars, augmented by machine learning, can extend our ability to assess offshore wind resources over broad areas. It offers insight into the trade-offs between spatial coverage and accuracy: even with some biases, the HF-radar-based approach captures the broad wind resource characteristics well, which could be invaluable for preliminary site assessments and identifying high-potential zones. Further refinements to the inversion model and additional validation (e.g., with more buoys or Lidar) will continue to improve confidence in this emerging remote-sensing approach for wind resource evaluation.

Acknowledgements

This work has been funded by the HF-EOLUS project (TED2021-129551B-I00), financed by MICIU/AEI /10.13039/501100011033 and by the European Union NextGenerationEU/PRTR - BDNS 598843 - Component 17 - Investment I3. Members of the Marine Research Centre (CIM) of the University of Vigo have participated in the development of this repository.

Disclaimer

This software is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages, or other liability, whether in an action of contract, tort, or otherwise, arising from, out of, or in connection with the software or the use or other dealings in the software.

References

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025a). HF-EOLUS HF-Radar Wind Inversion Toolkit for Artificial Neural Networks Training and Inference (v0.1.1) (v0.1.1). Zenodo. https://doi.org/10.5281/zenodo.17464519

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025b). HF-EOLUS. Task 2. HF-Radar Wind Inversion Models and Results for VILA and PRIO Stations. Zenodo. https://doi.org/10.5281/zenodo.17464583

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). Project HF‑EOLUS. Task 1. Puertos del Estado Vilano Buoy Data Bundle (GeoParquet + STAC) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17098037

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). HF EOLUS Wind Resource Toolkit (v0.1.0) (v0.1.0). Zenodo. https://doi.org/10.5281/zenodo.17591545

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). HF-EOLUS. Task 5. Wind Resource Estimation Results. Zenodo. https://doi.org/10.5281/zenodo.17594220