Subspace time series clustering of meteocean data to support ocean and coastal hydrodynamic modelling

Abstract

High-quality ocean and coastal circulation predictions strongly rely on the availability of accurate time series of main meteocean forcing, e.g. wind velocity, atmospheric air pressure. In this study, we formulated and tested on a real case a new approach for generating medium term time series of meteocean data from available reanalysis database. We used the fifth generation reanalysis (ERA5) of global climate and weather data. The methodology is based on the K-means clustering technique, which groups unlabeled data into different clusters. In particular, we implemented a subspace clustering method that includes an automatic weighting of the meteocean variables of interest. To test the performance, we apply the algorithm to the Pearl River Estuary (China). By comparing the proposed methodology with standard clustering analysis, our results suggest that the obtained meteocean scenarios (clusters) better reproduce the time trends of the main variables, in terms of typical indexes used for evaluating the goodness of clustering. Moreover, we showed how using climatological averages, commonly adopted for circulation and wave models, could lead to loosing the important local variability of the meteocean signals. The present approach could represent an advancement in time series clustering to be coupled with ocean and regional circulation models in ocean engineering applications.