| 摘要: |
| 构建了一种长短时记忆神经网络（LSTM）和全连接神经网络（FC）结合的臭氧（O3）预测模型（LSTM-FC），并考虑O3质量浓度的周期性变化规律，以珠三角为例，实现了对其进行高精度预测的目标。结果表明：（1）考虑周期性的LSTM-FC模型24 h预测结果的均方根误差（RMSE）为16.08 μg/m3，决定系数（R2）可达0.82，相比未考虑周期性的模型，精度提升了32.28%。（2）考虑周期性的LSTM-FC模型对O3质量浓度低值部分能够取得更精确的预测结果，对高值部分低估的现象改善效果显著。考虑周期性后，大于《环境空气质量标准》（GB 3095—2012）中O31 h平均质量浓度一、二级限值的预测结果均得到了一定改善，RMSE分别下降了18.71%和34.90%，R2分别提升了40.42%和134.04%。研究结果表明，考虑周期性的LSTM-FC模型在O3预测方面具有良好的拓展性和应用潜力。 |
| 关键词: 臭氧 预测模型 深度学习 周期性 珠三角 |
| DOI： |
| 分类号:X515 |
| 基金项目:国家自然科学基金项目（42201359）；广东省自然科学基金面上项目（2022A1515010492）；中山大学大学生创新训练计划项目（20212045） |
|
| Reach on a Deep Learning Approach Considering Periodicity for Ozone Prediction Models： A Case Study of the Pearl River Delta |
|
CHEN Dairong， CUI Yuxiang， SU Yuenong， WU Jingan， LI Tongwen
|
|
School of Geospatial Engineering and Science，Sun Yatsen University，Zhuhai， Guangdong 519082， China
|
| Abstract: |
| In this paper， an ozone prediction model which combines the Long Short Time Memory (LSTM) neural network and the Full Connection (FC) neural network (denoted as LSTM FC) is constructed， and the periodic variation rule of ozone concentration data is introduced to realize the high precision prediction of ozone concentration. Taking the Pearl River Delta as an example， the results show that： (1) the root mean square error (RMSE) of 24 hour prediction results of LSTM FC model that considers periodicity is 16.08μg/m3， the coefficient of determination (R2) can reach 0.82， which improves the accuracy by 32.28% compared with the model without considering periodicity. (2) Considering the periodicity， the LSTM FC model can achieve more accurate prediction results for the low value part. At the same time， it can significantly improve the phenomenon of high value underestimation. The RMSE of prediction results that greater than the national primary and secondary ambient air quality standards decreased by 18.71% and 34.90% respectively， and R2 increased by 40.42% and 134.04% respectively. The above results show that the proposed LSTM FC model possesses good expansibility and application potential in ozone prediction. |
| Key words: Ozone Prediction models Deep learning Periodicity Pearl River Delta |
