文章通过BP神经网络模型, 利用西沙站的实测潮位推算三亚站潮位, 研究用一地点的潮位资料去推算另一地点(异地)潮位的方法。文章比较了不同隐含层节点数和输入因子对潮位推算结果的影响, 采用预测时间(t)之前N个小时(t-N+1, …, t-1, t)西沙站的实测潮位数据作为输入因子, 输入因子数目在2~10之间, 隐含层分别采用节点数3、4、5、10和15建模, 分多种情况进行推算。结果显示, 对文中使用的特定情形, 隐含层为4个节点的效果最好, 隐含层为15个节点的效果最差; 输入层为2个节点的效果最好, 输入因子增多会使得推算效果变差。隐含层为4个节点、输入因子为t-1、t时刻潮位的仿真验证的结果最好, 推算值和实测值之间的相关系数为0.9901, 均方根误差为0.06m, 误差在-0.16~0.15m之间。结果表明, 如果两个地点的潮位具有物理上的关联, 通过BP神经网络模型, 用一地点的实测潮位推算另一地点潮位的方法是可行的。

Based on a BP neural network (NN) model, we investigated the possibility to derive the tide level at one site using the tide observation at a nearby site. As an example, the application of the neural network model to predict the tide level at Station Sanya using the observed data at Station Xisha is presented. The observed tide level at different hours ahead of the prediction time (t-N+1, …, t-1, t) are used as the input vector. Combinations of different numbers of input-layer nodes (2 to 10) and hidden-layer nodes (3, 4, 5, 10, 15) are used, and the outputs are compared with field data. For the specific case used in this study, four nodes in the hidden layer lead to the best prediction while 15 nodes performs the worst; two nodes in the input vector are the most suitable while more input nodes lead to degraded performance. The model setup with a hidden layer of four nodes and an input vector of two variables (t-1, t) has the best prediction accuracy in this case, with a correlation coefficient of 0.9901, a root-mean-square (RMS) error of 0.06 m, and a prediction error of -0.16~0.15 m (between prediction and observation). The proposed NN model is examined for its applicable to predict tides at a station using the observed tide data of a nearby station if they are physically coherent with similar tidal type.