| 基于多通道长短期记忆网络的PM2.5小时浓度预报 |
| |
| 引用本文： | 张鑫磊,张冬峰,刘伟,杨倩,郭媛媛,任玉欢,范志宣.基于多通道长短期记忆网络的PM2.5小时浓度预报[J].环境科学研究,2022,35(12):2685-2692. |
| |
| 作者姓名： | 张鑫磊 张冬峰 刘伟 杨倩 郭媛媛 任玉欢 范志宣 |
| |
| 作者单位： | 1.山西省气候中心，山西 太原 030006 |
| |
| 基金项目： | 山西省应用基础研究计划项目(No.201901D211572)；山西省应用基础研究计划项目(No.20210302124626)；山西省气象局青年基金项目(No.SXKQNTQ20217142) |
| |
| 摘 要： | PM2.5作为主要的大气污染物之一，严重影响空气质量和人体健康. 基于深度学习的PM2.5小时预报研究中，不同输入要素的历史时间序列对PM2.5预报结果的响应情况存在差异. 因此，基于太原市2019—2020年空气质量监测站、气象观测站的数据，提出一种多通道长短期记忆网络(Multi-Channels Long Short Term Memory，MULTI-LSTM)模型对PM2.5浓度进行预报. 首先使用独立的长短期记忆网络(LSTM)学习每个输入要素，然后将每个模型的学习结果进行融合，最终获得未来多小时的PM2.5浓度预报结果. 将单通道LSTM模型(BASE-LSTM)和LSTM扩展模型(LSTME)作为对照模型，与MULTI-LSTM模型的预报精度进行对比. 结果表明:不同观测窗与预报时效下，MULTI-LSTM模型在测试集上的预报精度明显高于其他2个对照模型. 其中，MULTI-LSTM模型在8 h观测窗和6 h预报时效组合下，均方根误差(RMSE)、平均绝对百分误差(MAPE)以及拟合指数(IA)分别为20.26 μg/m3、51%、0.91. 对未来逐6 h的预报中，观测窗宽度从8 h增至32 h，MULTI-LSTM模型的预报精度无明显变化，观测窗宽度为40和48 h时，RMSE比8 h观测窗下分别下降了2%和3%. 此外，增加LSTM层深度不会提升模型的预报精度. 研究显示，利用MULTI-LSTM模型作为PM2.5浓度小时预报模型，通过选取合适的观测窗宽度与气象要素，可获得精度较高的预报结果.
|
| 关 键 词： | 长短期记忆网络(LSTM) PM2.5浓度预报 时间序列 多通道 深度学习 |
| 收稿时间： | 2022-05-30 |
Hourly Concentration Prediction of PM2.5 Based on Multi-Channels Long Short Term Memory |
| |
| Affiliation: | 1.Shanxi Climate Center, Taiyuan 030006, China2.Jilin Institute of Meteorological Sciences, Changchun 130062, China |
| |
| Abstract: | As one of the major air pollutants, PM2.5 seriously affects air quality and human health. Hourly prediction of PM2.5 is important for fine control of PM2.5 and preventing cardiovascular and respiratory diseases related to PM2.5. In the research of PM2.5 hourly prediction based on deep learning, there are differences in the response of historical time series of different input factors to PM2.5 prediction results. Therefore, based on the data from air quality monitoring stations and meteorological observation stations in Taiyuan from 2019 to 2020, a multi-channels long short-term memory neural network (MULTI-LSTM) model was proposed. Firstly, each input feature was learned by an independent long short-term memory neural network (LSTM). Then, the learning results of each model were merged to obtain the PM2.5 concentration in the next several hours. Single-channel LSTM model (BASE-LSTM) and LSTM extended model (LSTME) were compared with MULTI-LSTM model. The results showed that in different observation windows and period validities, the MULTI-LSTM model was superior to the other models. And the RMSE, MAPE and IA of MULTI-LSTM model were 20.26 μg/m3, 51% and 0.91, under the 8 hours observation window and 6 hours period validity. When the period validity was 6 hours, and the width of the observation window was increased from 8 hours to 32 hours, the prediction accuracy of MULTI-LSTM model did not change significantly. The RMSE for the 40 hours and 48 hours observation windows decreased by 2% and 3%, respectively, compared to 8 hours. The comparison of the pollution process among different degrees, the MAPE of the MULTI-LSTM model in different pollution levels is lower than that of the other models. In the predicting of station, MULTI-LSTM model had the highest coincidence degree between the predicted value and the observed value. In addition, increasing the depth of LSTM layer did not improve the prediction accuracy to the model. Therefore, using the MULTI-LSTM model as the hourly prediction model of PM2.5 in Taiyuan. By selecting appropriate observation window and meteorological elements, high accuracy prediction results can be obtained. |
| |
| Keywords: | |
|
| 点击此处可从《环境科学研究》浏览原始摘要信息 |
|
点击此处可从《环境科学研究》下载全文 |
|
