基于季节性自回归积分滑动平均与深度学习长短期记忆神经网络的降水量预测
DOI:
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

P401

基金项目:

地面气象观测综合集成关键技术开发和应用(GYHY201306070)


Prediction and analysis of precipitation based on SARIMA and deep learning LSTM neural network
Author:
Affiliation:

Fund Project:

Development and application of key technologies for comprehensive integration of ground meteorological observations (GYHY201306070)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    降水量数据是一种非线性、非平稳的时序序列,传统的方法较难获取数据的变化规律,深度学习LSTM能较好地处理好多要素变量与降水量之间的非线性关系。本文利用扬州市区1960-2019年8种气象基本要素数据,采用传统SARIMA和深度学习LSTM神经网络方法对降水量数据进行预测比对,并着重分析了LSTM在不同类型不同输入与输出模式形态下的预测水平差异。结果表明:(1)传统的SARIMA模型中静态模式较动态模式能更好地反映出扬州市区月降水量数据变化趋势,且与实际值差距较小。动态模式容易造成误差累积或整体易呈现周期性稳态变化,实时性欠缺。(2)深度学习LSTM 多输入单输出动态预测模式下,完整周期的数据输入可以让神经网络更好地学习数据的变化规律。然而将多个周期数据作为一个输入单位,易造成模型过拟合。LSTM模型(look_back=12)对扬州市区月降水量预测精度优于传统的SARIMA模型,RMSE训练值低0.02。(3)LSTM多输入单输出动态模式(look_back=12)较LSTM多输入多输出静态模式,RMSE测试值低0.33,体现出该模式对扬州市区月降水量预测准确度更高。与此同时,M-LSTM多输入多输出静态模式预测准确度优于LSTM多输入多输出静态模式,RMSE测试值低0.19,反映出M-LSTM多输入多输出静态模式的优点。

    Abstract:

    The precipitation data is a non-linear and non-stationary time series. It is difficult to obtain the changing law of precipitation data with traditional methods. The deep learning LSTM (long short-term memory) network can better deal with the non-linear relationship between various meteorological variables and precipitation. The data of 8 basic meteorological elements in Yangzhou City from 1960 to 2019 was used to predict precipitation data by adopting traditional SARIMA (seasonal autoregressive integrated moving average) and deep learning LSTM network methods. At the same time, the difference in the prediction level of LSTM under different input and output modes was analyzed. It is concluded that: (1) The static model in the traditional SARIMA model can better reflect the change trend of monthly precipitation data in Yangzhou City than the dynamic model, and the difference between the predicted value and the actual value is smaller. The dynamic mode is easy to cause error accumulation or the whole is easy to show periodic steady-state changes, and lack of real-time. (2) Deep learning LSTM in the multiple-input-single-output dynamic prediction mode, the complete cycle of data input allows the neural network to better learn the changing laws of the data. However, if multiple periods of data are used as an input unit, it is easy to cause the model to overfit. The LSTM model (look_back=12) is better than the traditional SARIMA model in predicting monthly precipitation in Yangzhou City, and the RMSE training value is 0.02 lower. (3) The LSTM multiple-input single-output dynamic mode (look_back=12) is 0.33 lower than the LSTM multiple-input multiple-output static mode RMSE (Root Mean Squared Error) test value, which shows that the model has higher accuracy in predicting monthly precipitation in Yangzhou city. At the same time, the prediction accuracy of the M-LSTM (multi-category LSTM) multi-input-multi-output static mode is better than that of the LSTM multi-input-multi-output static mode, and the RMSE test value is 0.19 lower, reflecting the advantages of the M-LSTM multiple-input multiple-output static mode.

    参考文献
    相似文献
    引证文献
引用本文

张丽婷,李鹏飞,庞文静,等. 基于季节性自回归积分滑动平均与深度学习长短期记忆神经网络的降水量预测[J]. 科学技术与工程, 2022, 22(9): 3453-3463.
Zhang Liting, Li Pengfei, Pang Wenjing, et al. Prediction and analysis of precipitation based on SARIMA and deep learning LSTM neural network[J]. Science Technology and Engineering,2022,22(9):3453-3463.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-06-26
  • 最后修改日期:2022-02-16
  • 录用日期:2021-11-22
  • 在线发布日期: 2022-03-25
  • 出版日期:
×
律回春渐,新元肇启|《科学技术与工程》编辑部恭祝新岁!
亟待确认版面费归属稿件,敬请作者关注