Fine Grained Air Quality Prediction Based On Deep Learning

In recent years,the problem of air pollution caused by rapid economic development has become increasingly prominent,and the harm of urban air pollution to human health and climate has become increasingly serious.Therefore,effective and accurate prediction of air pollutant concentration has important guiding significance for the prevention of urban air pollution.In this paper,two methods were proposed for PM2 in Beijing,including an improved shallow machine learning model and a deep learning model.Because atmospheric prediction is a data-driven science,the traditional time series model is difficult to capture effective information in a large number of historical monitoring data,resulting in unsatisfactory results.With the rapid development of science and technology,machine learning,especially deep learning technology promote the development of artificial intelligence.In this paper,we built forecasting models based on Extreme Learning Machine（ELM）and Sequence-to-Sequence（Seq2Seq）model.（1）In the complex atmospheric environment,the multi-step prediction performance is unstable due to the noise and outliers caused by the complex and changeable environmental factors of sensors.In order to solve the above problems,the generalized relative entropy was introduced and compared with stack extreme learning machine（S-ELM）to construct a fast generalized relative entropy-based Stack Extreme Learning Machine（FGC-SELM）model.Experiments showed that it enhanced the robustness of the algorithm,improved the prediction accuracy,and had a certain practical significance.（2）Compared with deep learning model,traditional machine learning model has some limitations.In this paper,a variational air quality forecasting model based on attention mechanism was proposed.A recurrent neural network based on Variational Auto-Encoder（VAE）was introduced as the decoder of seq2seq model.Through a series of potential variables,the semantic information between outputs of different time steps could be captured,and then the self-attention mechanism was incorporated into the decoding process to alleviate the posterior failure problem of Bayesian model.The experimental results showed that it could effectively improve the prediction accuracy.Compared with the existing methods,it had certain advantages over the general deep learning model.Eventually,by utilizing the air quality data of Beijing,paper evaluated two prediction models with MAE and R~2.By comparing the results with other arts,it was proved that this study had obvious advantages in predicting air quality.Combined with the final prediction results,it provided a scientific and reasonable theoretical basis for the relevant departments to strengthen the prevention and control of air pollution,and offered a certain application value.