基于深度神经网络的GRACE等效水高及地表形变量预测

利用2002-04~2017-06 GRACE月时变重力场信息，反演得到三峡水库区域、长江流域和亚马逊流域的等效水高及地表垂直、水平形变时间序列。在LSTM(long short-term memory)网络基础上，通过堆叠LSTM结构以及在输出层中添加线性连接层来增加网络层数，构成深度LSTM神经网络对时间序列进行预测。引入注意力机制以提高模型对于序列长期特征的提取能力，并使用遗传算法筛选最佳网络层数和优化部分网络超参数。结果表明，在动态预测模式下，纳什系数NSE(Nash-Sutcliffe efficiency coefficient)最差为0.907 9，最好可达0.977 7，标准化的均方根误差R*(scaled root mean square error)最小为0.146 5，最大为0.297 5；在静态预测模式下，评价指标R*均低于0.062 2，NSE均大于0.99，表明模型性能优异。

Time Series Forecasting of Equivalent Water Height and Surface Displacements from GRACE Using Deep Neural Networks

Using GRACE(gravity recovery and climate experiment) monthly time-variable gravity field data from April 2002 to June 2017, we invert time series of equivalent water height and surface vertical and horizontal deformation in the area of Three Gorges Reservoir, Yangtze River basin and Amazon basin. A deep neural network model based on simple LSTM(long short-term memory) network is applied to predict time series data. The deep LSTM network could be extended deeper by stacking multiple LSTM hidden layers and adding linear layers in output layers. In addition, the attention mechanism is added to increase the ability of long-term characteristics extraction, and the genetic algorithm is used to select the best number of network layers and optimize some hyper parameters. In dynamic predicting mode, the NSE(Nash-Sutcliffe efficiency coefficient) is 0.907 9 at worst and 0.977 7 at best, and the values of R*(scaled root mean square error) are between 0.146 5 and 0.297 5. In static predicting mode, all of NSE values are better than 0.99, and the values of R* are less than 0.062 2, showing that the performance of deep LSTM network is very good.