基于深度学习提取时空信息的流域内库水位预测模型研究
Research on Model for Predicting Reservoir Water Level in Basin Based on Deep Learning and Extracting Spatiotemporal Information
周兰庭 1陈思思 1孙永明2
作者信息
- 1. 河海大学水利水电学院,江苏 南京 210098
- 2. 江苏省太湖水利规划设计研究院有限公司,江苏 苏州 215128
- 折叠
摘要
为了解决流域连通水库增多,库水位影响因素复杂且具有非平稳性,难以直接通过水文计算预测的问题,对流域水文站点日降雨序列进行分析,首先将时间序列经小波变换去噪,在此基础上采用最大信息系数(MIC)相关性分析筛选与日水位序列相关性,增加了输入时序降雨与预测水位相关的信息密度,并提出将强相关性序列输入引入 Attention机制的长短期记忆(LSTM)预测模型,提高 LSTM神经网络选择和提取序列特征的能力.以福建某流域站点实测日降雨序列为例进行试验,结果表明该方法的均方预测误差仅为 0.190 8,相比 LSTM模型有更高的预测精度,为水库水情调度及防洪减灾管理提供了决策依据.
Abstract
In order to solve the problem of the influence factors of the connected reservoir water level being complex and non-stationary,which is difficult to predict directly through hydrological calculation,the daily rainfall sequence of the hydrological station in a small basin was analyzed.Firstly,the time series was denoised by wavelet transform.On this basis,the maximal information coefficient(MIC)correlation analysis was used to screen the correlation of the daily water level series,which increased the information density related to the input rainfall series and predicted water level.A pre-diction model of the LSTM,which introduces strong correlation sequence input into attention mechanism,was proposed to improve the ability of LSTM neural network to select and extract sequence features.Taking the measured daily rainfall sequence of a small watershed in Fujian Province as an example,the results show that the prediction error of this method is only 0.190 8.Compared with the LSTM model,this method effectively improves the prediction accuracy,which pro-vides a decision basis for the operation of reservoir water situation and flood control and disaster reduction management.
关键词
库水位预测/相关性分析/小波变换/Attention机制/LSTMKey words
reservoir water level prediction/correlation analysis/wavelet transform/attention mechanism/LSTM引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据