Adaptive nonlinear state-space modelling for the prediction of daily mean PM10 concentrations

An adaptive nonlinear state space-based modelling system has been designed to predict daily mean concentrations of PM10 for Bordeaux metropolitan area. The nonlinear model structure is based on empirical relationships between the measured PM10 and other primary pollutants and meteorological variables. An Extended Kalman filter algorithm is used to estimate 1-day ahead prediction of the extended state, containing model parameters and daily mean PM10. A key characteristic of such a system is that its behaviour can be adapted to the short-term changes of air pollution and consequently the model can handle the time-evolving nature of the phenomena and does not need frequent adjustments. The method is applied to data from a monitoring site in Bordeaux (south France). Experimental results show that the model accurately predicts daily mean PM10. The application of the Extended Kalman filter explains about 70% of the variance with an absolute mean error less than 4.5 μg/m3. The approximate index of agreement value for the period covered is 0.90.