Clustering of Spatiotemporal Data with Applications to Wind Speeds

• Activity Number: 297
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Laura Tupper* and David S. Matteson and C. Lindsay Anderson
• Companies: Cornell University and Cornell University and Cornell University
• Keywords: time series ; clustering ; high-dimensional data ; spatiotemporal data ; depth statistics

Abstract

Our work is motivated by the need to find representative behaviors for wind speed across time and multiple locations. Short-term variability in wind speed presents a major challenge to its integration into the power grid, and the shape of the wind speed curve is as important as the overall level. To assess differences in such time series, we propose a functional distance measure, the band distance, extending the band depth of Lopez-Pintado and Romo (2009). This measure can be used to cluster observations without reliance on pointwise Euclidean distance, so as to emphasize the shape of time series, functional, or high-dimensional observations relative to other members of a dataset; it is invariant to monotonic transformations of the data, and performs well in the presence of heteroskedasticity. By combining this method with processing of the wind data, such as removal of seasonality or use of the time-frequency domain, we can go beyond mean-dependent standard clustering methods, such as k-means, to provide more shape-influenced cluster representatives. We also present an application using ECIS (electric cell-substrate impedance sensing) data to characterize different cell types.