Based on the theory that the high dimensional phase space of a forced dissipative nonlinear system will eventually evolve into a low dimensional attractor, while a great deal of small-scale high frequency patterns are dissipated, an approach of improving extended range forecast is put forward. In which a set of EOF bases holding up the climatic attractor is gained from historical weather data to remove small-scale high-frequency disturbances which distort the monthly mean field.Two numerical experiments of monthly mean forecast, for January and July, have been performed by using T42 global spectral model and FGGE data. The results show that the AC of monthly mean forecasting field is apparently increased by applying the approach put forward in this thesis, which means the approach has effect on enhancing the forecast skill.