Monte Carlo ensemble forecast method is employed to establish a surface ozone ensemble forecast system in Beijing. The ensemble forecast system is constituted by 50 forecast members with 154 model input parameters of Nesting Air Quality Prediction Model System (NAQPMS) perturbed. The forecast skill of the surface ozone ensemble forecast system is evaluated over three days’forecast (11－13 Aug 2008) during Beijing Olympics 2008. Analysis is performed on probabilistic forecast skill, deterministic forecast skill and on how to represent forecast uncertainty. Results indicate that, compared with single model, ensemble forecast significantly improves accuracy of ozone hourly concentration and daily maximum prediction. By selecting sub-ensemble with smaller error, the root mean square error of forecast is reduced by over 10%.