改进的遗传BP交叉训练算法及其在河流冰情预报中的应用

A modified genetic and BP alternately training algorithm and its application to river ice forecasting

本文在对传统的遗传算法(GA)及BP网络进行一系列改进的基础上,将遗传算法与BP网络进行了有机的结合,提出了一种交叉训练算法。该算法的核心在于将BP网络训练过程划分为若干子步,使遗传进化机制和BP网络训练机制步进交叉进行,来保证网络训练的成功率和快速向全局最优区域逼近的目的。把该算法应用到松花江依兰、佳木斯江段封河、开河日期预报中,取得了良好的效果。

This paper puts forward a novel alternately training algorithm that combines the genetic algorithm （GA） with the BP neural network and a series of improvements on the traditional algorithms. The key of this algorithm is to partition a BP training process into several steps, so that GA operation and BP training can be performed alternately step by step, aiming at a final success of the training and a fast convergence toward global optimization. Application of the algorithm to the Yilan and Jiamusi reaches of Songhuajiang River shows a high accuracy in forecasting the freeze-up and break-up dates.