基于模糊神经算法的区域地下水盐分动态预测  

Regional groundwater salinity dynamics forecasting based on neuro-fuzzy algorithm

作　　者：余世鹏[1,2] 杨劲松[1,2] 刘广明[1] 姚荣江[1,2] 王相平[1] 

机构地区：[1]中国科学院南京土壤研究所,南京210008 [2]中国科学院南京分院东台滩涂研究院,东台224200

出　　处：《农业工程学报》2014年第18期142-150,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金资助项目(41101518;41171181);江苏省产学研联合创新资助项目(BY2013062);江苏省自然科学基金资助项目(BK2011883)

摘　　要：为探讨前馈型人工神经网络BP-ANN(back propagation artificial neural network)和模糊神经NF(neuro-fuzzy)2种神经网络算法在区域地下水盐分动态预测中的应用过程与效果,首先通过经典统计分析确定区域地下水盐分动态的主要驱动因子以及可用的模型输入因子组合,采用"试错法"确定神经网络模型的最优结构,进而开展地下水盐分中长期动态的有效模拟预测。结果表明,在长江河口寅阳和大兴地区以降水动态为单输入的NF(5-gbellmf-160)和以降水与内河水盐分动态为双输入的NF(4-gaussmf-100)为最优预测模型。研究表明神经网络模型对地下水盐分动态的预测精度优于常规线性模型,其中,NF、BP-ANN、线性模型在寅阳测点的预测相关系数分别为0.565、0.445、0.261,在大兴测点的预测相关系数分别为0.886、0.784、0.543。与BP-ANN、线性模型相比,基于模糊神经算法的NF模型具有更好的误差纠错和仿真能力,在寅阳和大兴测点的预测误差分别降低了30%以上和50%以上。相关研究结果在区域水盐动态科学预警研究领域有较好地应用前景。The study conducted a detailed analysis of the modeling processes and performances of 2 types of different neural network models including back propagation artificial neural network （BP-ANN） and neuro-fuzzy （NF）, in the groundwater salinity dynamics forecasting. Firstly, the classical statistical analysis was used to determine the dominant driving factors of groundwater salinity dynamics and to reveal the available model inputs combinations. Then, the optimal neural network model structures were determined by the trial-and-error method and used to effectively forecast the mid-long term groundwater salinity dynamics. By our research, the idea of necessity in selecting the optimal NF model parameters of transfer functions, rule numbers and iteration steps was innovatively proposed, and the mechanism of differences involved in the model inputs for different groundwater salinity dynamics forecasts was demonstrated. At estuarine Yinyang site, the optimal NF forecast model structure was NF（5-gbellmf-160） with 1 input of the precipitation dynamics, which denotes the optimal rule numbers of 5, the bell type transfer function and the iteration steps of 160. The optimal BP-ANN forecast model structure was ANN（2-2-1）, which denotes 2 inputs of precipitation and river water EC dynamics, 2 hidden layers and 1 output. As for estuarine Daxing site, the optimal NF forecast model structure was NF（4-gaussmf-100） with 2 inputs of precipitation and inland water EC dynamics, which denotes the optimal rule numbers of 4, the gauss type transfer function and the iteration steps of 100. The optimal BP-ANN forecast model structure was BP-ANN（1-3-1）, which denotes 1 input of inland river EC dynamics, 3 hidden layers and 1 output. As the dominant groundwater recharge resource, the precipitation dynamics was the major impact factor on estuarine groundwater salinity dynamics. On the other hand, the groundwater salinity dynamics at Yinyang site was also affected by the high river water salinity, while at Daxing site was

关 键 词：水 盐分 土壤 地下水盐分动态 人工神经网络 模糊神经算法 最优模型参数 中长期预测 

分 类 号：S273.4[农业科学—农业水土工程]