---

作 者 ：卢志娟,朱玲,裴洪平,汪勇

期 刊 ：生态学报 2008年 28卷 10期 页码：4965~4973

关键词：小波分析；BP神经网络；Chl-a；短期预测；杭州西湖；

Keywords:wavelet analysis, BP neural network, chlorophyll-a, short-term prediction, the West Lake,

摘 要 ：小波神经网络是基于小波分析理论所构造的一种分层的、多分辨率的新型人工神经网络。选择合适的小波基和分解尺度对西湖水体Chl-a进行小波分析，将原序列分解成一个低频概貌分量和多个高频细节分量，再通过BP网络建立西湖叶绿素a浓度短期预测模型Ⅰ和模型Ⅱ。模型Ⅰ将小波分析去除高频细节信息后的低频概貌部分作为输入变量预测Chl-a含量；模型Ⅱ则对低频部分和高频部分分别进行预测，最后汇总各分网络输出得到最终结果。对确证集预测时，模型Ⅰ的平均误差为44%，模型Ⅱ仅为1.9%，且误差范围较模型Ⅰ小，表明模型Ⅱ具有较高的预测精度和稳定性。最后运用模型Ⅱ进行水质预测，预测值与实际值的平均相对误差为6.4%，并选取3号点（中山码头）进行模型的泛化，平均相对误差为6.9%，取得了较理想的预测效果，说明小波神经网络能成功预测西湖水体中Chl-a含量的短期变化趋势，为西湖水质管理提供科学依据。

Abstract:Wavelet neural network is a new kind of hierarchical and multiresolution artificial neural network which based on wavelet analysis theory. In this paper, we choose an appropriate wavelet base and decompose scale to analysis chlorophyll-a of the West Lake. We divide the original sequence into a low frequency and several high frequency parts, then establish model Ⅰ and model Ⅱ for short-term prediction of chlorophyll-a concentration in the West Lake through BP neural networks. The model Ⅰ uses low frequency part only as input for network to forecast the content of chlorophyll-a, while the model Ⅱ uses the low frequency and the high frequency part as inputs, then summarizes the outputs to get the final product. Comparing with the two models, we can see the average error of model Ⅱ is smaller than of model Ⅰ,and the scope of error is also narrow. That means the precision and stability of model Ⅱ are higher than of the model Ⅰ. Finally, we forecast the water quality with the model Ⅱ. That shows the average relative error between predictive value and actual value is 6.4%. By selecting the third pot (Zhongshan dock) to generalize the model Ⅱ, which enable the average error is 6.9%. The result indicates that wavelet neural network can successfully forecast the content of Chlorophyll-a in the West Lake, so can provide scientific guidance for the West Lake management.

---