| 基于神经网络和数值模型的重点区域PM2.5预报比较分析 |
| 摘要点击 3565 全文点击 1173 投稿时间：2021-05-07 修订日期：2021-07-11 |
| 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 BP神经网络 NAQPMS模型 人工订正 重点区域 PM2.5预报 比较分析 |
| 英文关键词 BP neural network NAQPMS model artificial correction key regions PM2.5 forecast comparative analysis |
|
| 中文摘要 |
| 应用BP神经网络法建立京津冀及周边城市、汾渭平原、苏皖鲁豫交界地区和长三角地区等重点区域95个城市PM2.5预报模型，对2020年秋冬季上述地区城市开展未来7 d的PM2.5预测预报，并对比同期业务化运行的数值模型预报结果和各城市人工订正后预报结果，对3方法预报效果进行分析评估.结果表明：①4区域神经网络法模型性能短期预报相对较好，中长期有所降低，对4区域均有一定的系统性高估，苏皖鲁豫交界地区系统性偏差最小，长三角地区偏差最显著.数值模型区域预报水平较神经网络有所降低，各评价指标总体低于神经网络，对辖区城市间预报效果较神经网络差异更大.②神经网络、数值模型和人工订正方法对4区域PM2.5浓度预报准确率普遍较低，平均不足50%，准确水平总体呈：神经网络 > 人工订正 > 数值模型.3方法分指数级别范围准确率均大幅提升，4区域1~4 d平均准确率均在65%以上，神经网络模型和人工订正水平相近，总体高于数值模型.③在预报中度及以上污染级别日时，数值模型在京津冀及周边城市、苏皖鲁豫交界地区和长三角地区效果均较为理想，汾渭平原最差.神经网络模型对京津冀及周边城市、苏皖鲁豫交界地区和汾渭平原短期预报效果较好，长三角地区较差.人工订正结果总体在中度污染级别时预报效果相对较好，重度及以上预报效果和神经网络模型相近. |
| 英文摘要 |
| The PM2.5 forecast models of 95 cities in Beijing-Tianjin-Hebei and its surrounding cities (BTH); the Fenwei Plain (FWP); the border area of Jiangsu, Anhui, Shandong, and Henan (JASH); and the Yangtze River Delta (YRD) regions were established using BP neural network models, and the forecast was carried out for the next seven days in the autumn and winter in 2020. By comparing the forecast results of the BP neural network models, numerical model, and artificial correction, the PM2.5 forecast effects of the three methods were analyzed and evaluated. The results showed:① The performance of the short-term forecast based on the BP neural network was relatively good but was reduced in the medium and long term and systematically overestimated in four regions. The numerical model effects were lower than those of the BP neural network models. ② The accuracy rates of the PM2.5 forecast concentration by the three methods were generally low in the four regions, with an average of less than 50%, and the accuracy values in order from high to low were the BP neural network models, artificial correction, and the numerical model. The accuracy rates of IAQI levels of PM2.5 were significantly improved by the three methods, and the averages were above 65% in the first four days. The effects of the BP neural network models and artificial correction were similar, which were generally higher than those of the numerical model. ③ The numerical model had good effects in the BTH, JASH, and YRD regions, whereas it was the worst when forecasting moderately and above-polluted days in the FWP region. The BP neural network model had a good performance when forecasting short-term PM2.5 in the BTH, JASH, and FWP regions, whereas it was poor in the YRD region. In general, the performance of artificial correction was relatively good when forecasting moderate-level days and was close to the BP neural network model when forecasting heavily polluted days. |
|
|
|
