A component time-series model for SO2 data: Forecasting, interpretation and modification

Abstract A time-series forecasting method is developed to enable advance warning of smog in winter. A component model for the time series of SO 2 concentration essentially using a recursive Kalman algorithm is constructed on the basis of spectral analysis. It is found that the smog episodes with low frequencies and time-dependent power spectra are solely represented by the trend component. This component is therefore investigated in the phase space, where it exhibits a typical trajectory. For forecasting, one part of the data is used to establish the parameters and another part is used to test the extrapolation. The extrapolation and interpolation behaviour of the Kalman filter used is investigated. The trend component is found not to agree with the behaviour of the trajectory in the phase space. A modified method is proposed to extrapolate the time-dependent spectrum of the trend component, namely local harmonic approximation. This method is tested and compared with linear extrapolation and is found to provide a generalisation, producing closer correspondence between the concentration values predicted and those actually observed.