地球信息科学学报 ›› 2020, Vol. 22 ›› Issue (8): 1617-1629.doi: 10.12082/dqxxkx.2020.190378

• 地球信息科学理论与方法 • 上一篇    下一篇

基于LSTM神经网络的青藏高原月降水量预测

刘新(), 赵宁, 郭金运*(), 郭斌   

  1. 山东科技大学测绘科学与工程学院,青岛 266590
  • 收稿日期:2019-07-16 修回日期:2019-10-01 出版日期:2020-08-25 发布日期:2020-10-25
  • 通讯作者: 郭金运 E-mail:xinliu1969@126.com;jinyunguo1@126.com
  • 作者简介:刘 新(1969— ),女,山东肥城人,博士,副教授,主要从事空间数据挖掘和机器学习等研究。E-mail:xinliu1969@126.com
  • 基金资助:
    国家自然科学基金项目(41774001);国家科技基础性工作专项(2015FY310200)

Prediction of Monthly Precipitation over the Tibetan Plateau based on LSTM Neural Network

LIU Xin(), ZHAO Ning, GUO Jinyun*(), GUO Bin   

  1. College of Geomatics, Shandong University of Science and Technology, Qingdao 266590, China
  • Received:2019-07-16 Revised:2019-10-01 Online:2020-08-25 Published:2020-10-25
  • Contact: GUO Jinyun E-mail:xinliu1969@126.com;jinyunguo1@126.com
  • Supported by:
    National Natural Science Foundation of China(41774001);The Basic Science and Technology Project of China(2015FY310200)

摘要:

青藏高原的降水量预测不仅为该地区水资源合理规划利用提供依据,同时对中国及周边国家气候变化研究有着重要的意义。论文利用1990—2016年青藏高原降水量数据,采用长短期记忆神经网络(LSTM)对青藏高原月降水量进行预测,主要包括:① 使用青藏高原86个测站1990—2013年的月降水资料,预测各个测站2014—2016年的月降水量,并与传统的RNN、NAR、SSA和ARIMA预测模型相比,平均决定系数R2分别提高了0.07、0.15、0.13和0.36,均方根误差(RMSE)和平均绝对误差(MAE)表现更低;② 分析了降水量预测精度的空间分布特征,将各模型的R2在青藏高原地区内插值,分析R2的空间分布特征,发现所有模型降雨稀少的干旱地区和降雨多的湿润地区R2较低,在气候稳定、降水规律性明显的地区R2较高,且LSTM模型R2≥0.6的空间范围远大于传统模型;③ 分析了不同预测长度对各模型预测精度的影响,发现所有模型会随着预测长度增加而预测精度降低,但在不同的预测长度下LSTM预测的RMSE值都低于其他模型。

关键词: 长短期记忆网络, 降水量, 预测, 循环神经网络, 青藏高原, 时间序列, 机器学习

Abstract:

Precipitation prediction on the Qinghai-Tibet Plateau not only provides a basis for rational planning and utilization of water resources, but also has significance for climate change research in China and neighboring countries. In this paper, the Long Short Term Memory neural network (LSTM) was used to predict the monthly precipitation over the Qinghai-Tibet Plateau using data from 1990 to 2016. Firstly, the monthly precipitation data of 86 stations in the Qinghai-Tibet Plateau from 1990 to 2013 were used to predict the monthly precipitation of each station from 2014 to 2016. Comparing with the traditional RNN, NAR, SSA, and ARIMA prediction models, LSTM increased the average coefficient of determination (R2) by 0.07, 0.15, 0.13, and 0.36, respectively. Simultaneously, LSTM had lower Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). Among them, the observation of station 56106 showed that the LSTM model predicted the period more accurately with less displacement deviation, and that the prediction of the valley between July and September was more accurate with R2 reaching 0.87. Secondly, the spatial distribution characteristics of precipitation prediction accuracy were analyzed. The R2of each model was interpolated in the Qinghai-Tibet Plateau, and the spatial distribution characteristics of R2were analyzed. All the drought areas with rare rainfall and the wet areas with heavy rainfall were of lower R2, while the areas with stable climate and obvious precipitation were of higher R2. Areas of R2 over 0.6 were much larger when using the LSTM model than the traditional model. Finally, influence of different prediction lengths on the prediction accuracy was analyzed for each model. All models showed decreased prediction accuracy as the prediction length increased, yet the RMSE values predicted by LSTM were lower than by other models with the varying prediction lengths.

Key words: LSTM neural network, precipitation, prediction, RNN neural network, Qinghai-Tibet Plateau, time series, machine learning