CAJ | 학술논문

传统BP(back propagation)算法在实际应用中具有网络结构参数和学习训练参数难以确定、泛化能力差、训练学习易陷入局部极小点等问题.该文在传统BP算法的基础上,提出一种改进算法,在训练过程中能自动确定各种参数,并避免陷入局部极小点,提高网络的泛化能力.利用2003-2005年5-9月中国国家气象中心T213的数值预报产品,通过动力诊断得出反映降水的物理量,然后从中挑选出与降水关系较好的25个因子,连同中国国家气象中心T213模式、日本气象厅业务模式和德国气象局业务模式相应的降水量预报结果作为预报因子.采用改进的BP算法建立江淮流域68个站24 h降水(08:00-08:00,北京时)3个等级(降水量≥0.1 mm,降水量≥10 mm,降水量≥25 mm)的预报模型.通过对2006-2007年5-9月68个站试报结果表明:改进BP算法对降水预报的TS评分大大高于传统BP算法,也高于几种模式的降水预报结果,同时,改进算法使降水预报的平均空报率、漏报率明显降低.
전통BP(back propagation)산법재실제응용중구유망락결구삼수화학습훈련삼수난이학정、범화능력차、훈련학습역함입국부겁소점등문제.해문재전통BP산법적기출상,제출일충개진산법,재훈련과정중능자동학정각충삼수,병피면함입국부겁소점,제고망락적범화능력.이용2003-2005년5-9월중국국가기상중심T213적수치예보산품,통과동력진단득출반영강수적물리량,연후종중도선출여강수관계교호적25개인자,련동중국국가기상중심T213모식、일본기상청업무모식화덕국기상국업무모식상응적강수량예보결과작위예보인자.채용개진적BP산법건립강회류역68개참24 h강수(08:00-08:00,북경시)3개등급(강수량≥0.1 mm,강수량≥10 mm,강수량≥25 mm)적예보모형.통과대2006-2007년5-9월68개참시보결과표명:개진BP산법대강수예보적TS평분대대고우전통BP산법,야고우궤충모식적강수예보결과,동시,개진산법사강수예보적평균공보솔、루보솔명현강저.
Objective forecast of precipitation is difficult because of its complex nonlinear characteristics.In order to enhance the ability of forecasting precipitation,artificial neural network (ANN) method is applied in numerical weather products interpretation.Among different types of ANN,the back propagation (BP) neural network is the most popular and influential one.However,traditional BP algorithm has some limi-tations such as the difficulties in determining network structure and the learning parameters,poor general-ization ability and possibility of misleading to local minimum in learning process,etc.To resolve these problems,an improved algorithm is proposed. Based on T213 numerical forecast products of National Meteorological Center from May to September during 2003-2005,25 factors are selected in terms of dynamic diagnostic analysis and statistical methods. The precipitation forecasts of operational global models from China National Meteorological Center,Japan Meteorological Agency and German Meteorological Administration are studied.Using the reformative BP algorithm,three grades forecast(≥0.1 mm,≥10.0 mm,≥25.0 mm)models are built to forecast 24-hour precipitation of 68 stations over Jiang-Huai Basin.During the training process,precipitation samples are randomly divided into two kinds according to a certain proportion,training samples and testing sam-pies.They are used to train the network and to check the error of output respectively so that all parame-ters are confirmed.By repeating training and learning of network,an optimal network model is obtained. The optimized forecast model is used to forecast precipitation of different grades,times and stations,from May to September during 2006-2007.The forecasting results of improved BP algorithm are compared with those of tradition BP algorithm and numerical models outputs.The average threat score (TS) of im-proved BP algorithm is the highest;the average false alarm rate (FAR) and missing alarm rate (MAR) of improved BP algorithm are much lower than the others.So the improved BP algorithm is superior and it indicates a potential for more accurate precipitation forecasting.