Selecting daily ECMWF (European Center for MediumRange Weather Forecasts) numerical forecast grid field data at 20:00 BT from March 1,2003 to December 31, 2008, the forecast factor database that can reflect the local weather dynamic characteristics is constructed by using such methods as difference method, weather diagnosis and factor combination. And a BP neural network prediction model of the daily highest and minimum temperatures of various months and stations is established by first, roughly checking factors with PRESS (prediction square sum) criteria, second, checking again factors with stepwise regression, and finally, careful checking factors with optimal subset regression, thus the 1-5 day test forecast of maximum and minimum temperatures is done. The result of operational model trial shows that, the BP neural network prediction model has a strong nonlinear processing capability, and can better reflect the changes of daily extreme temperature, thus the average forecast accuracy of 1-5 day maximum and minimum temperatures reaches to higher levels. It is sensitive to warming and cooling processes. The trend and range forecasts of warming and cooling are more correct. It provides an important technical support to the precise town temperature forecast within 1-5 days. Meanwhile, it is a good idea and method of the application of the ECMWF numerical forecast products to temperature forecast.