基于BP神经网络的鄱阳湖水位模拟

考虑到鄱阳湖水位受流域五河与长江来水等多因素的共同作用而表现出高度非线性响应，采用典型的三层BPNN神经网络模型来模拟鄱阳湖水位与其主控因子之间的响应关系。分别将湖口、星子、都昌、棠荫和康山水位作为目标变量进行BPNN模型构建和适用性评估。结果显示：综合考虑流域五河及长江来水（汉口或九江）的BPNN水位模型，空间站点水位模拟精度（R2和Ens）可达090以上，各站点的均方根误差（RMSE）变化范围约050～10 m，若忽略长江来水的影响作用，仅将流域五河来水作为湖泊水位的主控影响因子，模型训练期与测试期的纳希效率系数（Ens）和确定性系数（R2）显著降低，且低于050，均方根误差（RMSE）也明显增大（124～288 m），意味着综合考虑流域五河与长江来水是获取结构合理、精度保证的鄱阳湖水位模型的重要前提。同时建议针对鄱阳湖湖盆变化对水位的影响，尽可能选择一致性较好的长序列数据集来训练和测试BPNN模型。所构建的BPNN神经网络模型可进一步结合流域水文模型，用来预测气候变化与人类活动下流域径流变化对湖泊水位的潜在影响，也可作为一种有效的模型工具来回答当前鄱阳湖一些备受关注的热点问题，如定量区分流域五河与长江来水对湖泊洪枯水位的贡献分量，为湖泊洪涝灾害的防治和对策制定提供科学依据

USING BP NEURAL NETWORKS FOR WATER LEVEL SIMULATION IN POYANG LAKE

Poyang Lake is the largest freshwater lake in China, and has an internationally recognized wetland system. It receives inflows predominantly from five major rivers (i.e. the Ganjiang, Fuhe, Xinjiang, Raohe and Xiushui Rivers) within its drainage catchment. The Ganjiang, Fuhe and Xinjiang Rivers contribute approximately 89% of the lakes inflow from its drainage catchment area, and the remaining 11% is mainly made up of inflows from the Raohe and Xiushui Rivers. Poyang Lake is connected to the Yangtze River through a narrow channel at Hukou at its northern end, and the Yangtze River plays a complementary role in controlling outflows.Because the Poyang Lake has a non linear response to the river discharges from its drainage basin and the Yangtze River, the highly complexity and nonlinear characteristics determine that the three layer back propagation neural network (BPNN) has the ability to simulate the lake water level responses. In this study, the water level time series of the Hukou, Xingzi, Duchang, Tangyin and Kangshan were used as the target variables for the BPNN construction under different model scenarios. Results indicate that both discharges from the catchment rivers and the Yangtze River are considered, thedetermination coefficients R2 and Nash Sutcliffe efficiency Ens for all gauging stations can reach 090 and the Root Mean Square Error RMSE is in range of 050-10 m during the BPNN training and testing phases, while the simulated results reveal that the Ens and R2 are lower than 050, and RMSE is significantly increased in range of 124-288 m under the condition of neglected the Yangtze River discharges. The BPNN can be used to combine the catchment hydrological models, which can provide an alternative tool for predicting the lake water levels in response to catchment river discharges under climate and land use changes. The BPNN also can be used as an effective modeling tool to solve some hot issues in the Poyang Lake, such as how to quantitatively distinguish the individual contributions of catchment rivers and the Yangtze River to the flood and low water levels. At the same time, the model users should select long time series dataset with best data consistency to train and test the BPNN model