基于递归小波神经网络的江苏城镇夏季最高气温预报预警技术

Summer maximum temperature forecast and warning in Jiangsu cities and towns based on recurrent wavelet neural network

针对目前数值天气预报产品释用方法上所存在的释用因子固化,无法应对特殊转折性天气的问题,应用一种基于动态因子检验的递归小波神经网络(Recurrent Wavelet Neural Network,RWNN)对江苏城镇夏季最高气温进行释用。该方法可以自动选取气象要素且无需建立回归方程,具有泛用性好、灵活性高的特点。使用该方法基于T639的2017-2018年6-8月资料建立了江苏省南京、徐州、射阳、常州、苏州5地的最高气温预报预警模型。实验结果表明:南京、徐州、射阳3地模型的TT2和HSS35评分较反向传播神经网络方法分别平均提高了9个百分点和0.15,同时较卡尔曼滤波方法分别平均提高了17个百分点和0.2。

To solve the problem that the meteorological factors used in the interpretation of numerical forecast products are fixed and unsuitable for special turning of weather phenomenon,a recurrent wavelet neural network(RWNN)method based on dynamic factors was adopted to interpret the summer maximum temperature in Jiangsu cities and towns.This method could select meteorological factors automatically and with out establishing statistical equations,so could be widely used.Based on the data of five cities and towns of Jiangsu Province including Nanjing,Xuzhou,Sheyang,Changzhou,Suzhou during June-August of 2017-2018 from T639 numerical forecast products,the maximum temperature forecast and warning models were established with the method.Experimental results show that models’TT2 scores have increased by 9%and HSS35 scores have increased 0.15 compared to back propagation neural network,and models’TT2 scores have increased by 17%and HSS35 scores have increased 0.2 compared to Kalman filter on average in Nanjing,Xuzhou and Sheyang.