GA-BP神经网络模型在流域面雨量预报的应用研究

GA-BP ANN MODEL FOR RIVER CATCHMENT PRECIPITATION FORECAST

通过采用遗传算法优化网络初始权重的方法,将遗传算法(GA)和前馈误差反传播(BP)算法有机地结合,优势互补,并应用于流域面雨量预报。以广东省东北部的滨江流域为试验区域,以1995～2001年气象探空资料为基础,利用最优子集回归技术进行预报因子筛选,建立了流域面雨量预报的GA-BP神经网络模型,取得了满意的结果。试验表明:(1)6小时流域面雨量预报神经网络的优化结构是7-7-1,转移函数的组合方式为tansig-线性函数。(2)训练算法为Levenberg-Marquardt算法(LM)。(3)遗传算法具有快速学习网络权重的能力,对BP网络易陷于局部极小点。(4)利用GA-BP神经网络模型对未来6小时流域面雨量的预报精度明显高于其它统计方法,证明了这种方法的有效性和可靠性。

The method is taken to join the genetic algorithm （GA） and BP algorithm together and supplement mutually by optimizing the initial weights of ANN with GA, and some applications have been made in the Binjiang River catchment for precipitation forecast. The ANN model by GA has been established in which forecasting variables are selected by optimizing the subclass regression technique on the base of radiosonde data （from 1995 to 2001） and the optimized ANN model for 6 hours precipitation of Binjiang River catchment has been obtained. The optimized ANN structure is 7-7-1 and its transfer function is tansig-pureline and training functions is Levenberg-Marquardt（LM）. The genetic algorithm can speed up the learning process of network weights and solve the local searching problem of BP network. The experiment result shows that this method can enhance the forecast precision of 6-hours precipitation compared with other statistical methods, and its effectiveness and the reliability have been proved.