基于深度学习的PM_(2.5)浓度预报  

Prediction of PM_(2.5)Concentration Based on Deep Learning

作　　者：吕秋明 莫欣岳 李欢 Lv Qiuming;Mo Xinyue;Li Huan(School of Cyberspace Security/School of Cryptology,Hainan University,Haikou,China)

机构地区：[1]海南大学网络空间安全学院(密码学院),海南海口

出　　处：《科学技术创新》2023年第22期77-80,共4页Scientific and Technological Innovation

基　　金：教育部产学合作协同育人项目(220902070162538);中国高等教育学会高等教育科学研究规划课题(22LH0409);海南省自然科学基金(623RC455,623RC457);海南大学科研启动基金项目(KYQD(ZR)-22096,KYQD(ZR)-22097);海南大学教育教学改革研究项目(hdjy2364)。

摘　　要：空气污染严重威胁生态环境和人体健康,开展及时准确的空气质量预报至关重要。基于卷积神经网络(CNN)、双向长短期记忆网络(BiLSTM)和注意力机制(Attention)等多种深度学习方法,建立了深度学习组合模型CNN-BiLSTM-Attention,并针对北京市PM_(2.5)小时浓度开展预报。实验结果表明:超参数调节能有效提高深度学习模型性能;CNN-BiLSTM-Attention可以准确预报PM_(2.5)浓度变化;与基准模型LSTM相比,提出的组合模型引入了多种深度学习算法的优点,在MAE、MSE和R2统计指标上分别提升了80%、94%和0.3%。Air pollution poses a serious threat to the ecological environment and human health,and timely and accurate air quality prediction is crucial.Based on convolutional neural network(CNN),bidirectional Long short-term memory network(BiLSTM),attention mechanism and other deep learning methods,a deep learning combination model CNN-BiLSTM-Attention is established,and the hourly concentration of PM_(2.5)in Beijing is predicted.The experimental results show that hyperparameter adjustment can effectively improve the performance of deep learning models;CNN-BiLSTM-Attention can accurately predict PM_(2.5)concentration;compared with the benchmark model LSTM,the proposed model introduces the advantages of multiple deep learning algorithms,and improves MAE,MSE,and R2 by 80%,94%,and 0.3%,respectively.

关 键 词：空气污染预报 深度学习 双向长短期记忆网络 注意力机制 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]