Improvement of Ensemble Technique Using Spectral Analysis and Decomposition of Air Pollution Data

Abstract

The current study proposes a novel approach for the multi-model ensemble to be applied in air pollution forecasting. The methodology is based on decomposition of air pollution time series on different components (short-term, daily fluctuations, synoptic scale, etc.) and calibration of the ensemble for each of these components independently taking into account the performance of individual predictors. Therefore, the same model may have a different contribution for the ensemble at high and low frequency fluctuations. The Kolmogorov-Zurbenko (KZ) low-pass filter is used for the time series decomposition. The Fourier analysis is implemented to determine the contribution of different frequencies to the data variance allowing better understanding of the model performance and to define the ensemble weights. The methodology was tested using a group of four different air quality models that were applied over mainland Portugal for the 2006 year, and for main pollutants like O3 and PM10. The approach implemented in this work was compared with one of the most used ensemble technique showing clear advantages. © Springer Science+Business Media Dordrecht 2014.