最优子集神经网络高温预报模型

 利用2003-2007年6～9月ECMWF格点场资料，使用差分法、天气诊断、因子组合等方法构造出能反映本地天气动力学特征的预报因子库，采用press准则初选因子，尝试用最优子集方法进行神经网络夏季6～9月≥35℃高温预报模型的建模方法研究。2008年7月预报系统投入业务应用，检验证明所构造的神经网络高温预报模型具有更好的拟合和预报效果，为神经网络在灾害性天气预报的应用研究提供了新的思路和方法。

High temperature prediction modeling of neural network using the optimal subset

The changes in synoptic process are characterized by the obvious non-linear variation features,while the method of neural network is provided with strong abilities of disposing the non-linear problems.≥35 ℃ extreme high temperature over northwest arid region of China is severe weather in summer,it can cause difficult drinking water of man and domestic animal,aggravate drought,and has adverse influence on agriculture,industry and man’s health.Based on the daily data of ECMWF grid field at 20∶00-20∶00 from June to September during 2003-2007,the coverage is between 90-110°E and 35-45°N,grid is 2.5°×2.5°,height level is from 850 hPa to 500 hPa,the methods including difference,weather diagnosis and factors combination were used in this paper.Factors are selected by model output products which are closely related to temperatures or variations of low atmospheric condition,considering the continuity of the temperature changes,the factors were initially elected according to the press criterion.Optimal subset,stepwise regression and BP neural cell network method are attempted to form statistic forecast model for ≥35 ℃ extreme high temperature over the east of Hexi Corridor in Gansu province.Every month has its independent forecast equation.The result shows that factors selected by optimal subset method have better fitting,prediction ability and stability,whose calculated amount are so small that are easy to integrate in operational system,the selected factors are focused on near middle and low atmosphere which are T and T-Td at 850 hPa,relative humidity at 700 hPa.Meantime,forecasted factor was dealt with non-linearly which can reflect the non-linear relations between itself and other model output factors.The forecast system was verified by historical data during 2003-2007,its regional ≥35 ℃ high temperature fitting percent was 81.0% within 24 hours,and the forecast accuracy rate was 77.0% within 24 hours from July to September in 2008.The constructed model of the optimum subset’s high temperature forecast has been used in the actual forecast since summer of 2008,and is found to have the better fitting and high prediction accuracy.Thus,it provides a new train of thought and method for the research of the neural network prediction on the disastrous weather forecast.