1 as Jornadas de Engenharia Hidrográfica Lisboa, 21-22 de Junho de 2010 Correlating Wave Hindcast and Buoy data with Artificial Neural Networks Almeida, L.P. (1)., Vousdoukas, M.V. (1), Ferreira, P.M. (2,3), Ruano, A.E. (3), Dodet, G. (4), Loureiro, C. (1), Ferreira, Ó. (1), Taborda, R. (4). (1) University of Algarve, CIMA – Centre of Marine and Environmental Research (2) Algarve STP – Algarve Science & Technology Park, Faro, Portugal (3) University of Algarve, CSI – Centre for Intelligent Systems (4) University of Lisbon – Faculty of Sciences ABSTRACT: This work presents results from the use of Artificial Neural Networks (ANN) to improve wave models hindcasting capacity off the South coast of Portugal. Comparison of the original model results with field measurements showed significant non linear deviations. To compensate for such deviations, a three-layer Multilayer Perceptron (MLP – a type of an ANN) was trained, using the Levenberg-Marquardt method, to improve the fit between the hindcast (generated by WW3) and Faro buoy data in an effort to reconstruct missing data from the wave buoy time series. The results obtained so far are very positive; with the training with annual datasets showing better results than the training with the entire dataset, while both improved significantly the fitting of the raw model results. Further improvements are expected by trying different ANN types, by searching for optimised ANN input-output structure, and by performing sub-set selection on the data sets. KEYWORDS: Artificial Neural Networks, Hindcast wave model, wave data. 1. INTRODUCTION In-situ deployed instruments, such as buoys, are invaluable sources for continuous long-term oceanographic data acquisition, which is fundamental for a variety of research and operational applications. Given the rough and unfavourable conditions in the ocean, data gaps are a common consequence of instrument loss, malfunctioning, or delayed maintenance and data collection, being a major problem for data analysis. As an example Figure 1 characterize the data gaps (more than 3h without data) present on a wave time series obtained from the Faro Directional Wave Buoy (Portuguese Hydrographical Institute - IH), between 2000 and 2006. Figure 1. Number and length of gaps present in a wave time series obtained from Faro Wave buoy (IH). Periods without data can last from three hours to several days or weeks and simple interpolation methods are usually not an acceptable approach. One solution for these cases is to fill the gaps with results obtained by a wave generation model, e.g. WAVEWATCH III™ (Tolman, 2009). However wave models are highly-dependent on the quality and resolution of the wind forcing (Cavaleri and Bertotti, 2006; Ponde de Léon and Guedes Soares, 2008) and model results often show significant deviations from the real data at sheltered locations where local winds play an important role on the wave climate (Figure 2). Figure 2. Linear correlation between Hs Faro buoy and WW3, for the year 2004. One way to solve this non linear problem is through the use of artificial neural networks (ANN), since these are a modelling tool that can be used to model complex relationships between inputs and outputs. There are already in the bibliography some examples of the applications of ANN's in ocean engineering like to reconstruct wave time series using information from neighbourhoods buoys (Deo and Kumar, 2000; Puca et al., 2001; Londhe and Panchang, 2007;Medina and Serrano-Hidalgo, 2004) or for wave forecasting (Tsai et all., 2002; Makarynskyy , 2004; Makarynskyy et al., 2005;) using historical data from buoy for training.