| 基于KZ滤波和LSTM的上海市O3预测模型 |
| 摘要点击 1434 全文点击 282 投稿时间：2023-11-16 修订日期：2024-01-26 |
| 查看HTML全文
查看全文 查看/发表评论 下载PDF阅读器 |
| 中文关键词 O3预测 KZ滤波 LSTM 最大最小冗余 灰色关联性 支持向量回归（SVR） |
| 英文关键词 O3 prediction KZ filter LSTM maximal relevance and minimal redundancy gray correlation support vector regression（SVR） |
|
| 中文摘要 |
| 针对臭氧长时间序列预测精度及气象特征选择问题,提出通过Kolmogorov Zurbenko（KZ）滤波分解臭氧（O3）原始序列,利用支持向量回归（SVR）改进的最大最小冗余（mRMR）方法筛选气象特征,之后采用长短期记忆网络（LSTM）实现对上海2023年5~8月静安监测站（城市）、浦东川沙（城郊）和淀山湖（郊区）的O3预测.结果表明,特征筛选得到对O3基准和短期分量的最优组合包括气压、温度、湿度、边界层高度和风向.基于LSTM模型对特征筛选后的不同时间分量进行预测,静安站、浦东川沙站和淀山湖站R2分别为0.86、0.83和0.85,RMSE分别为18.26、18.74和20.02 μg·m-3,表明通过分解原始O3序列能够提高预测精度,特征筛选后的组合能够保持模型的预测性能. |
| 英文摘要 |
| In this study, a Kolmogorov-Zurbenko （KZ） filter was proposed to decompose the original ozone （O3） sequence to improve the accuracy of ozone long-term series prediction and select relevant meteorological features. Furthermore, the enhanced maximal minimal redundancy （mRMR） feature selection technique was combined with the support vector regression （SVR） approach to select the most illuminating meteorological features. Subsequently, from May to August 2023, during high ozone concentration periods, a long short-term memory network （LSTM） was utilized to assess and predict high ozone concentration periods at the monitoring stations of Jingan （urban area）, Pudong-Chuansha （suburban area）, and Dianshan Lake （suburban area） in Shanghai. The results showed that pressure, temperature, humidity, boundary layer height, and wind direction were the best combinations of O3 baseline and short-term components, as chosen by feature screening. The R2 values for Jingan Station, Pudong-Chuansha Station, and Dianshan Lake Station were 0.86, 0.83, and 0.85, respectively. The RMSE values were 18.26, 18.74, and 20.02 μg·m-3, respectively. These findings suggest that decomposing the original O3 sequence improved the prediction accuracy of ozone concentrations. Additionally, as indicated by the R2 and RMSE values found for every monitoring station, feature screening preserved the model’s predictive performance. |
|
|
|
