Identifying predictive multi-dimensional time series motifs: an application to severe weather prediction