A Kalman-filter bias correction method applied to deterministic, ensemble averaged and probabilistic forecasts of surface ozone
Article first published online: 11 FEB 2008
Journal compilation © 2008 Blackwell Munksgaard
Volume 60, Issue 2, pages 238–249, April 2008
How to Cite
DELLE MONACHE, L., WILCZAK, J., MCKEEN, S., GRELL, G., PAGOWSKI, M., PECKHAM, S., STULL, R., MCHENRY, J. and MCQUEEN, J. (2008), A Kalman-filter bias correction method applied to deterministic, ensemble averaged and probabilistic forecasts of surface ozone. Tellus B, 60: 238–249. doi: 10.1111/j.1600-0889.2007.00332.x
- Issue published online: 11 FEB 2008
- Article first published online: 11 FEB 2008
- (Manuscript received 26 June 2007; in final form 12 November 2007)
Kalman filtering (KF) is used to estimate systematic errors in surface ozone forecasts. The KF updates its estimate of future ozone-concentration bias using past forecasts and observations. The optimum filter parameter is estimated via sensitivity analysis. KF performance is tested for deterministic, ensemble-averaged and probabilistic forecasts. Eight simulations were run for 56 d during summer 2004 over northeastern USA and southern Canada, with 358 ozone surface stations.
KF improves forecasts of ozone-concentration magnitude (measured by root mean square error) and the ability to predict rare events (measured by the critical success index), for deterministic and ensemble-averaged forecasts. It improves the 24-h maximum ozone-concentration prediction (measured by the unpaired peak prediction accuracy), and improves the linear dependency and timing of forecasted and observed ozone concentration peaks (measured by a lead/lag correlation). KF also improves the predictive skill of probabilistic forecasts of concentration greater than thresholds of 10–50 ppbv, but degrades it for thresholds of 70–90 ppbv. KF reduces probabilistic forecast bias. The combination of KF and ensemble averaging presents a significant improvement for real-time ozone forecasting because KF reduces systematic errors while ensemble-averaging reduces random errors. When combined, they produce the best overall ozone forecast.