Cluster Analysis Resolution of Diurnal Climatological Wind Pattern Modes Utilizing K-Means – A Case Study with Boston, MA. Data (Logan International Airport, 1945-2019)

Climatological wind variability is an important meteorological element to be considered in planning and decision-making activities in which winds are a factor on some level. Conventional analytical tools such as wind rose diagrams and resultant wind statistics generally have a single hour focus, either for specific individual hours, or individual-hour composites of differing hours. Of interest also should be insights derived from analyses of contiguous-hour wind patterns. In the same manner as there are favored individual hourly directions and related speeds, there should be preferred, adjacent hour-to-hour patterns, or “modes. To this end, a multivariate treatment has been developed with good operational results for a number of stations in Southern California (Fisk, 2012, 2013, 2016, 2017), and the purpose here is to demonstrate its application on Boston, MA observational data, namely Logan International Airport recordings from 1945 into June 2019 (more than 20000 observations, 100% intact on a midnight-to-midnight individual hourly basis).

The focus is on statistical expressions of climatological wind “progressions” (for example, 24-element arrays of resultant wind statistics covering midnight to midnight), each array describing and contrasting statistically dissimilar idealized patterns of hour-to-hour climatological resultant winds. To create these arrays, raw individual hour wind observations (direction and speed) are decomposed into their u and v components, and then entered into a K-means cluster analysis (for example, 24 pairs of u and v components for a midnight-to-midnight treatment - a cluster analysis in 48-D space) . Resulting centroid values, by cluster, are then recombined into arrays of 24 resultant wind directions and speeds for each hour of the day. The K-Means treatment (i.e., number of clusters and cluster memberships) is enhanced by an optimizing data mining training/testing capability - the V-Fold Cross Validation Algorithm.

A given cluster’s hourly resultant direction and speed results allow inferences on diurnal climatological changes in wind directions and speeds from midnight-to-midnight. Percentage frequencies of the cluster patterns’ can be calculated by calendar month, and an appealing property is that, since a cluster analysis by its very nature assembles individual cases of the same statistical character into the same cluster, this frequently results in relatively high magnitude “constancy” statistics (mean vector wind speed divided by mean scalar wind speed times 100) among the 24-individual hour elements in summary array. Thus, the hourly mean resultants are can be frequently interpretable as reasonable approximations of average wind directions and mean scalar speeds, an intuitively satisfying outcome. Exceptions would be hours of the day which correspond to preferred diurnal changeover times in directions (seabreeze to landbreeze, or vice-versa). Interpretations of these, however, are no less straightforward.

Application of the methodology on the Boston data resolved five modes, and these results are presented in both graphical and tabular forms along with supporting explanations.