Abstract: Based on temperature forecast of operational middle-range model of China, operational model of German meteorological administration, operational model of Japan meteorological agency and temperature observations of China, a temperature consensus forecast system is developed through method of artificial neural network. Product of the system is station temperature forecast of China with 3-hour interval within 72 hours.Forecast modes of summer half year and winter half year are established separately. In order to include most recent impact of data, the process of developing forecast mode runs once a week under the condition of absorbing new data as much as possible.The system has been running stably from 1st of January in 2004. Testing of forecast result from January to May in 2004 indicates that consensus forecast is better than single model forecast. Absolute forecast error of consensus is less than 3.0 ℃ within 72 hours, and it has no systematical error. That means consensus can provide objective forecast support for forecast people. And it also indicates that artificial neural network is a kind of effective method to temperature consensus forecast.Consensus forecast error is different according to different area. It is bigger over Xinjiang and Xizang and smaller over south China and the reaches of the Ynagtze. The cause of this phenomenon is possibly that temperature variability is bigger over Xizang than that over the reaches of the Yangtze. Forecast error of consensus has obvious daily variation. It is always bigger during daytime than in night. On average, the consensus forecast also has forecast ability for temperature changing process with much more argument through contrasting between observation and corresponding forecast result of June in 2004 at partial stations. But the ability of forecast temperature with big argument is poor over the Tibetan area.In order to investigate single forecast impact to consensus forecast, different consensus schemes are developed with forecast results of different schemes checked. Contrasting between different schemes shows that every single forecast with good impacts is important for consensus forecasts results.