|
StormZ tracks spatially coherent groups of atmospheric risk signals through time. Daily groups are linked into tracks according to spatial proximity and temporal continuity, producing evolving risk structures rather than treating individual observations as independent points.
We investigated whether these tracked structures exhibit recurring temporal patterns in three StormZ variables: energy anomaly (E_anom), convection amplitude (conv_amp), and convection-direction change (rotation_score). Sixteen-day sequences were independently normalised by variable and compared using constrained multivariate Dynamic Time Warping (DTW).
This allows similar sequences to be recognised even where their component changes occur several days apart rather than simultaneously.
Importantly, predicted days to genesis (track_d2g_center) was not used in identifying the patterns. It remains available for subsequent testing of whether independently detected patterns have a relationship with storm-development timing.
For the currently available North Atlantic dataset, 997 continuous 16-day patterns were identified. Across 496,506 pairwise comparisons, DTW distances were p05 0.528, median 0.847, and p95 1.130.
After excluding temporally overlapping windows belonging to the same risk track, 483,012 comparisons remained and the distribution changed only marginally: p05 0.537, median 0.848, p95 1.130.
This indicates that the observed North Atlantic temporal similarity is not primarily an artefact of overlapping rolling windows.
For the longer currently processed South Pacific combined-domain dataset, 239 continuous 16-day patterns were identified from 8,880 track-state observations and 2,716 tracks.
After removing overlapping windows, 27,767 comparisons remained, with p05 0.739, median 1.011, and p95 1.286. Only 2.37% of pairwise comparisons were excluded because of temporal overlap.
The two basins should not presently be interpreted as directly comparable dynamical populations. Their geography, available historical periods and StormZ candidate populations differ substantially.
In particular, the combined southern domain encompasses several geographically distinct regimes, whereas the North Atlantic candidature domain is considerably more bounded.
Spatially tracked StormZ risk structures contain detectable recurring multivariate temporal behaviour that survives removal of overlapping-window artefacts.
The next research question is whether particular recurring sequences occur preferentially before tropical-cyclone genesis. Because genesis timing was withheld from the DTW pattern discovery, this relationship can subsequently be tested rather than assumed.
StormZ presently has only a fraction of its approximately decade-long underlying atmospheric archive processed through the computationally intensive candidate-selection and risk-track pipeline. Full historical reconstruction is therefore a proposed validation experiment, not part of these preliminary findings.