Details

Autor(en) / Beteiligte

• Peres, D.J.
• Iuppa, C.
• Cavallaro, L.
• Cancelliere, A.
• Foti, E.

Titel

Significant wave height record extension by neural networks and reanalysis wind data

Ist Teil von

• Ocean modelling (Oxford), 2015-10, Vol.94, p.128-140

Ort / Verlag

Elsevier Ltd

Erscheinungsjahr

2015

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• •Significant wave height is modeled by neural nets and NCEP/NCAR Reanalysis 1 data.•The proposed modeling technique is advantageous to extend observed records.•u and v 10 m winds are used to derive six-hourly 1948-present wave height series.•The effect of the number of wind points and of the wind-wave time lag is analyzed.•The reconstruction is generally of better quality than the NOAA CFSR one. Accuracy of wave climate assessment is related to the length of available observed records of sea state variables of interest (significant wave height, mean direction, mean period, etc.). Data availability may be increased by record extension methods. In the paper, we investigate the use of artificial neural networks (ANNs) fed with reanalysis wind data to extend an observed time series of significant wave heights. In particular, six-hourly 10 m a.s.l. u− and v−wind speed data of the NCEP/NCAR Reanalysis I (NRA1) project are used to perform an extension of observed significant wave height series back to 1948 at the same time resolution. Wind for input is considered at several NRA1 grid-points and at several time lags as well, and the influence of the distance of input points and of the number of lags is analyzed to derive best-performing models, conceptually taking into account wind fetch and duration. Applications are conducted for buoys of the Italian Sea Monitoring Network of different climatic features, for which more than 15 years of observations are available. Results of the ANNs are compared to those of state-of-the-art wave reanalyses NOAA WAVEWATCH III/CFSR and ERA-Interim, and indicate that model performs slightly better than the former, which in turn outperforms the latter. The computational times for model training on a common workstation are typically of few hours, so the proposed method is potentially appealing to engineering practice.