改进的人工神经网络模型在水文序列预测中的应用研究

Improved artificial neural network model for hydrologic time series prediction

目前人工神经网络(ANN)技术在水文序列模拟预测中有较多应用.由于径流时间序列往往呈现出复杂的变化过程,直接使用径流序列建立的ANN单变量预测模型大多精度较差而难以满足需要.本文拟从两个方面进行改进,以提高ANN径流量预测模型的精度.首先,根据径流序列的变化规律,滤去序列中的季节性变化趋势,并采用局部多项式拟合求残差方法消除局部波动成分后得到新序列作为ANN模型训练样本,结果表明由此训练得到的ANN模型比未作处理的样本训练得到的ANN模型预测精度明显改善.另一方面,通过消除趋势波动分析(DFA)方法检测径流序列的分形特征并估算其标度区间;选取不同数量的训练样本进行训练得到多个ANN模型,确定模型预测精度最高时的训练样本数,并与标度区间比较,结果表明在标度区间内选取训练样本数可明显提高ANN模型预测精度.这对ANN模型训练样本数的选取有指导意义.

The generation of runoff was affected by several factors.Runoff series always contain information of multiple factors,which make the variation of runoff complex.People usually model the regular pattern of the hydrologic time series by applying a univariate artificial neural network（ANN） model to simulate the variational process.However,it is often unacceptable because of the low prediction precision.We analyze the variation of the hydrologic time series and deseasonalize and detrend it before inputting it for ANN training data.The results show that the ANN that uses processed training data excelled the ANN that uses original data for training data.Meanwhile,we detect the fractal characteristic of the hydrologic time series and estimate the range of invariant scale employing the detrend fluctuation analysis（DFA） method.We choose different numbers of training data and get corresponding ANN models.The results of the experiments indicate that we have more chances to obtain the best ANN model when select a training data with a number with in the scale range.