改进IGA-GP方法在流域需水预测中的应用--以黄河流域为例

Application of improved IGA-GP method in valley water forecast:a case study in Yellow River

遗传算法是借鉴生物学的自然选择和遗传进化机制,通过作用于染色体上的基因寻找优良的染色体,求到待解问题最优解的一个近似解。针对遗传算法易陷入局部最优解的不足,采用改进的遗传算法(IGA)与遗传编程(GP)相结合,建立有广泛搜索能力和很强的局部精化能力的IGA-GP算法,将该算法应用于BP神经网络的优化,优化的原理是运用GP自动生成BP正向计算过程中隐含层到输出层的函数关系式,此函数关系式可随着训练样本的变化而自动调整,使得新加入的样本影响已学习完的样本的问题得到最终解决,因而可进一步增强BP网络的自适应能力和抗干扰能力;同时用IGA代替BP的反向传播算法,加快收敛速度和克服BP易陷入局部最优的不足,从而实现了对BP神经网络的优化。并在此基础上建立了黄河流域需水预测模型,误差分析结果显示,IGA-GP算法预测结果最大误差为2.894%,最小误差为0.088%,满足误差精度不大于5%的要求,而没有优化的BP神经网络拟合结果最大误差为8.314%,最小误差为0.832%,不满足预测精度要求,证明该IGA-GP模型具有较高的预测精度。鉴于此,通过IGA-GP的算法对黄河流域2010、2020及2030水平年工业、农业、生活及生态需水量进行了预测。

Genetic Algorithm is to find an approximate approach to the optimal solution to an open problem with biont＇ s natural selection and the genetic evolution mechanism through seeking excellent chromosomes by genes which is acting on chromosomes. Due to the shortcomings of genetic algorithms （ GA）, which is prone to lose in the local optimal solution, this paper applied the algorithms combined improving genetic algorithms （IGA） with genetic programming （ GP）, which owns the virtues of better searching capability in comprehensive and local aspect than that of GA, to optimizing BP neural network in order to solve the shortcomings of being prone to lose in local optimal solution in the course of searching optimal solution of BP. The principle of optimization is generating functional relation from hidden layer to output layer in forward computational process automatically, which could adjust automatically with the change of training samples, and the influence of new added samples to the studied ones could be solved, therefore auto-adapted ability and anti-jamming ability of BP network could be further enhanced. Meanwhile using IGA to replace the back propagation algorithm of BP speeds up convergence and overcome defects of BP easily running into local optimum, thus it realizes the optimization of BP neural network. And then a model of water-requirement prediction in Yellow River Basin has been obtained on the basis of the method. The result of error analysis demonstrates that the maximum error of IGA-GP algorithm forecasting result is 2. 894% , and the minimum error is 0. 088% , which satisfies the requirement of erroneous precision should not be bigger than 5% ; while the maximal error of fitting results using BP neural network without optimization is 8. 314% , and the minimum error is 0. 832% , which is not agreeable with the requirement of erroneous precision. It proves that IGA-GP model has higher estimation precision. This paper has carried on the forecast of industry, agriculture, life and the ecology water demand to the Yellow River valley in 2010, 2020 and 2030 median water years with the IGA-GP algorithm.