针对现阶段水质监测中存在的水质变化响应滞后问题, 提出了采用灰色预测法、人工神经网络( BP 神经网络、 径向基神经网络、广义回归神经网络) 以及两者组合的方法对水质动态预测进行研究。以太湖流域嘉兴斜路港监测 断面为例, 并依据后验差检验比值 c 及小概率精度 p 对模型预测效果进行了分析。结果表明, 对年内预测, 通过广 义回归神经网络的动态预测值平均相对误差为 01 61% , 后验差检验比值小于 01 65, 小误差概率大于 01 7; 采用灰色 结合广义回归神经网络的方法对水质 pH 值进行预测, 平均相对误差仅有 01 85% , 后验差检验比值小于 01 65, 小误 差概率等于 1。研究结果还表明, 对年际预测, 灰色结合 BP 神经网络和灰色结合径向基函数神经网络的动态预测 值平均相对误差分别为 01 57% 和 01 80% , 其后验差比值都小于 01 5, 小概率误差都为 01 9, 大于 01 8。

In this paper, the gr ey theor y, a rtificial neural netw or k ( back2 pr opagation neura l netwo rk, radial basis function neural netw or k, and g ener alized regr ession neural net wo rk) , and the combinat ion of these two metho ds w as pr oposed to st udy the dy2 namic predictio n o f water quality. Taking the Xielug ang in Jiax in as an example, the model prediction effect w as analy zed based on the posterior difference test r atio ( c) and small pro bability accuracy ( p ) . T he r esults show ed t hat within a predict ion y ear, the av erag e relat ive err or of the dynamic predictio n value of the g ener alized r egr essio n neura l netw ork was 01 61% , and the c was less than 01 65, w hile the p w as g reater than 01 7, respectively . The r esults ex hibited that the predictio n value using the combinat ion of the gr ey theo ry and g ener alized r eg ressio n neur al netw or k, the av erag ed relative err or w as 01 85% , the c< 01 65, and the p= 11 0, r espectiv ely. The inter2y ea r pr edict ion based on the combinatio n o f the gr ey theor y with BP neura l netw or k and radial basis function neura l netw or k, the av eraged relat ive err or w as 01 57% and 01 80, respectively , and the ratio o f posterior er2 r or was less than 01 5, and the small pro bability err or w as 01 9, but g reater t han 01 8.

宁波市教育科学规划重点课题( 2019YZD010)