用于洞察多时间尺度水文过程的深度学习模型

Deep learning model for insight into multi time scale hydrological process

在流量预测工作中,搭建了由单元、输入门、输出门和遗忘门组成的循环神经网络长短期记忆(RNN LSTM)模型,能学习长期时间依赖和捕获非线性关系,用来预测每日、每周和每月时间尺度的河流流量。利用2006—2014年唐河流域倒马关水文站的多个气候和水文要素数据集,可以对深度学习模型进行训练和验证。通过设置多个时序步长,探讨记忆单元存储信息的长短对不同时间尺度的影响。研究发现,RNN LSTM模型在日尺度预测中表现出较好的预测性能,数据集的粗粒度特性是影响周尺度和月尺度预测性能的关键。

In the work of flow prediction, a long-term and short-term memory RNN LSTM model consisting of unit, input gate, output gate and forgetting gate is built, which can learn long-term time dependence and capture nonlinear relationship to predict the river flow at daily,weekly and monthly time scales. The deep learning model can be trained and verified by using the data sets of multiple climatic and hydrological elements of the Daomaguan hydrologic station in the Tanghe river basin from 2006 to 2014. By setting multi time steps, the influence of the length of information stored in memory cells on different time scales was discussed. It is found that the RNN LSTM model has good prediction performance in daily scale prediction, and the coarse granularity of the data set is the key to affect the prediction performance of weekly and monthly scales.