Abstract:At present, national prediction system for dekad and monthly extreme h igh temperature has not been undertaken. Here, a national forecast system for ex treme high temperature was developed. It is a powerful platform that regularly i ssued three kinds of predicting products at website with three methods of dynami cal monthly forecast model, statistical downscaling and physical analogue using percentile values at 90%, 95% and 99% as relative thresholds and 35℃or 38℃ as absolute thresholds. The products consist of probability prediction, the numb er of day exceeding 35℃ or 38℃ involving 7 periods: 1-10 day, 11-20 d ay, 21-30 day, 1-30 day, 6-35 day and 31-40 day. The spatial distribution of correlatio n coefficient between hindcast and observation shows that high value areas are d ifferent among these methods. So it is hard to say which method is better than o thers. Thus three predicted results should be considered as well as distribution of correlation coefficient before final extreme high temperature prediction. Th e predictions by these methods for July, 2007 were generally close to the observ ation. But the result by statistical downscaling method is better, and that by D ERF is worse relatively.