Pattern Recognition and Cluster Analysis of Wind and its Structure-Embedded Association with Other Meteorological Characteristics

104

As interest in sailing has increased in Taiwan, it has become essential to understand wind patterns in order to better organize sailing events. Penghu began to organize sail regatta in 1999 and extended the scale of their regatta from the national level to the international level. Knowledge of both wind speed and wind direction is equally critical to successful regatta management. In this dissertation, wind speed and wind direction observed in 2011 from the Penghu data buoy are used as an example. A simple computational method was developed to summarize 24 hours of observation into a minimum ellipse with only five parameters. This summarized daily ellipse reduces the dimensions of the problem efficiently but still retains hourly information. We demonstrate herein that this minimum ellipse computational method can be applied in at least three ways: (1) tracking the wind pattern trajectory of a typhoon, (2) providing a simple method to distinguish land-and-sea breeze days, and (3) visually exhibiting the global wind pattern for the planning of sailing events. Based on the results of the minimum ellipse, daily wind patterns can be summarized, and cluster analysis on a daily basis can be performed. An important clustering technique, called a data cloud geometry-tree (DCG-tree) is introduced in this dissertation. The DCG-tree clustering method provides better quantification of the multi-scale geometric structures of the data under consideration than the standard Hierarchical Clustering method. This property was verified from real data by clustering daily wind patterns in Penghu. The data mechanics method is another important concept which provides a tree structure-embedded linkage between the daily wind clusters and other meteorological covariates which were also observed from the data buoy in Penghu. The DCG-tree was used again to build the hierarchical structure of the relationships within the covariates which retain the information from daily wind pattern clusters. With the tree structure of these covariates, data mechanics was used to define the pairwise daily similarity from the point of view of the covariates. We then coupled the daily wind cluster with the daily covariate clusters. The coupling results indicated significant associations between these two types of meteorological characteristics. Taking into consideration predictions from the covariates to the wind patterns, the relationship built by the DCG-tree and data mechanics exhibited the same performance as that of the recently popular decision tree method. However, it provided more insight into the system dynamics and avoided the usual overfitting criticism lodged toward decision trees.