基于离散小波变换与模糊神经算法的河口日水位预测

Estuarine daily river stage predicting based on discrete wavelet transform and neuro-fuzzy algorithm

为了对受特殊复杂环境因子干扰的河口日水位中长期动态实现有效预测,将离散小波变换（DWT）和模糊神经（NF）算法结合,构建长江河口日水位中长期动态的联合预测模型DWT-NF。引入DWT方法对原始水位信号序列进行分解和减噪,筛选出最优DWT分解因子组合TD（D3＋D4＋D8）序列作为预测模型的输入。引入NF算法构建水位预测子模型,从滞后1～5天的TD序列中筛选出最优输入组合,确定河口日水位预测的最优DWT-NF模型结构,其中三条港水位预测的最优联合模型结构以滞后1～3天的TD序列为输入、以高斯函数和43分别作为NF子模型第一层的隶属度函数及其规则数,青龙港水位预测的最优联合模型结构以滞后1～4天的TD序列为输入、以钟型函数和24分别作为NF子模型第一层的隶属度函数及其规则数。综合比较表明,DWT-NF模型的预测效果显著优于其他联合模型及常规模型,尤其对河口未来中长期水位变化趋势的预测效果显著。

To effectively forecast estuarine daily river stage, which is influenced by complicated .environmental factors because of estuarine special geographic location, discrete wavelet transform （DWT） and neuro-fuzzy algorithm （NF） were combined to construct a hybrid DWT-NF model for mid-long term forecast of daily fiver stage in the Yantze estuary. This hybrid model uses DWT for decomposion of the original stage signals and filtering its jamming noises out. DWT revealed an optimum combination of .decomposition factors TD（D3＋D4＋Ds）, so we selected different TD serial combinations at a lag of 1 day to 5 days as inputs of the NF sub-model, and constructed an optimum WT-NF model for estuarine daily stage by training and testing different NF model structures. In forecasting the daily river stage in Santiao Port, the best NF model inputs with three nodes of TD serials at 1 day to 3 days lag were transferred into the first model layer using a Gauss function of rule number 43; in Qinglong Port, the best inputs with TD serials at 1 .day to 4 days lag transferred using a Bell function of rule number 24. Comparison with other hybrid or traditional models shows that the hybrid DWT-NF model has a much better performance, especially prominent in effectively forecasting detailed fluctuation trends of estuarine daily fiver stage dynamics.