Investigation of short-range cedar pollen forecasting

Abstract

Pollen forecasting is of increasing interest as a way to help the general public avoid contact with allergy-inducing pollen. It was recently reported that the dynamics underlying pollen concentration series is very similar to that of low-dimensional deterministic chaos, thus opening up new avenues of development in local pollen forecasting. Our analysis of hourly cedar pollen series for two seasons showed evidence of a small degree of determinism underlying the pollen time-series dynamics. However, we could not confirm that our pollen series was generated by a low-dimensional chaotic system. The nearest-neighbor method using local constant prediction applied to hourly pollen forecasting with a 1-h lead time was effective for small to medium pollen variations, but failed to reproduce large and intermittent pollen bursts. The performance of the nearest-neighbor model was significantly improved by applying a nonlinear filter to the source dataset. Standard time-series techniques such as neural networks did not improve upon these results. The difficulty in fully characterizing and accurately forecasting the pollen series was thought to originate in the nonstationarity of the series and in the large and intermittent pollen bursts that were found to have no apparent time structure. Thus the dynamics of hourly pollen series is probably not strongly tied to a low-dimensional chaotic system.