Acta Scientiarum Naturalium Universitatis Pekinensis ›› 2021, Vol. 57 ›› Issue (4): 645-652.DOI: 10.13209/j.0479-8023.2021.050

Previous Articles     Next Articles

Prediction of PM2.5 Daily Concentration of Guangzhou Based on Neural Network Algorithms

LI Zequn1, WEI Jun1,2,3,†   

  1. 1. School of Atmospheric Sciences, Sun Yat-sen University, Zhuhai 519082 2. Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai 519082  3. School of Marine Sciences, Guangxi University, Nanning 530004
  • Received:2020-06-03 Revised:2020-12-25 Online:2021-07-20 Published:2021-07-20
  • Contact: WEI Jun, E-mail: weijun5(at)mail.sysu.edu.cn

利用人工智能神经网络预测广州市PM2.5日浓度

李泽群1,  韦骏1,2,3,†   

  1. 1. 中山大学大气科学学院, 珠海 519082 2. 南方海洋科学与工程广东省实验室(珠海), 珠海 519082 3. 广西大学海洋学院, 南宁 530004
  • 通讯作者: 韦骏, E-mail: weijun5(at)mail.sysu.edu.cn
  • 基金资助:
    广东省重点领域研发计划(2020B1111020003)、国家自然科学基金(41976007, 91958101)和广西壮族自治区特聘专家专项经费(2018B08) 资助

Abstract:

Autoregressive integrated moving average (ARIMA) model, back propagation (BP) neutral network and long short-term memory (LSTM) are used to predict the daily concentration of PM2.5 in 2019 in Guangzhou city of China from 2015 to 2019. The effect of ensemble empirical mode decomposition (EEMD), temporal resolution on model prediction is explored in this paper. The results show that EEMD is able to improve significantly the prediction ability of the model on the low-frequency part of PM2.5 sequence. Increased temporal resolution can improve the prediction accuracy, with more input data. Since PM2.5 (t-1) is used as the input data, the model can only predict PM2.5 for 1 day in advance. To increase the prediction time window, we adopt a rolling forecast method, using PM2.5 (t) prediction value as the input data for PM2.5 (t+1). The result shows that the rolling forecast method allows the model to forecast PM2.5 (t+n) with a comparable MAE compared to the experiment without the rolling forecast method. In this paper, the ARIMA model (the time accuracy of input data is 6 hours) has the best prediction accuracy, and the minimum MAE value can reach 6.478. 

Key words:

摘要:

利用差分整合移动平均自回归模型(ARIMA)、后向传播神经网络(BP)以及长短期记忆神经网络(LSTM), 对广州市2015—2019年的PM2.5浓度数据进行训练和预报, 研究集合经验模态(EEMD)分解和时间分辨率对不同模型预报准确性的影响。结果表明, EEMD分解可以显著地提升低频分量的预报效果; 提高输入数据的时间分辨率可以提升预报效果, 尤其在ARIMA自回归模型预报中较为明显, 用神经网络进行预报时需要考虑输入数据量增加带来模型复杂度增加的问题。由于模型使用前一天(t -1)的PM2.5作为输入数据, 即只能预报t+1天的PM2.5值。为增加模型的预报时效, 采用滚动预报的方式对模型进行优化, 能够显著地提升预报时效, 实现对t+n天的连续预报, 且预报误差与后报结果相当。将时间精度为6 h的数据作为输入, 用ARIMA模型进行预报的效果最好, 最小MAE值为6.478。

关键词: 广州市, PM2.5, 整合移动平均自回归模型(ARIMA), 后向传播神经网络(BP), 长短期记忆神经网络 (LSTM), 集合经验模态分解(EEMD)