Study On Air Pollutant Concentration Prediction Method Based On Sequence Model

In recent years,with the accelerating industrialization,the air pollution problem has become more and more serious and has become an urgent problem in various countries.In addition to reducing the emission of pollutants at the source,many major cities have established air quality observation stations to monitor the air quality in real time.It is also important to forecast air quality in order to warn of extreme hazy weather and to guide people’s outdoor activities and travel planning.However,the current air quality prediction problem suffers from cross-interference of multiple data sources and error accumulation,and the most important concern of long-term air quality prediction and extreme weather warning is still challenging.To address the above problems,the authors propose two deep learning sequence models based on attention mechanism,and the main work and innovations of the paper are described as follows.(1)An encoder connection model based on temporal attention mechanism is proposed.In order to capture the correlation between multivariate sequences,the authors use two different encoders in the encoder stage to enhance the ability of learning features,and in addition,they introduce a temporal attention mechanism in the decoder stage to capture the long time step dependence of the sequences themselves.The method provides an effective solution to the single-step time prediction problem in air quality prediction.(2)A deep learning model(EvaNet)based on the extreme value attention mechanism is further proposed.The extreme value attention is innovatively introduced in the encoder part to enhance attention to the key part of the sequence of extreme values,combined with the grouping of enhanced feature capture between multiple sources of data obtained from data preprocessing and the encoding of the extreme value sequence that assists in calculating the attention score,and the temporal attention mechanism is followed in the decoding stage to capture the dependence of long time steps.The approach achieves significant results in the long-term air quality prediction problem.The above model is evaluated on two real-world city(Fuzhou and Beijing Shunyi District)datasets for PM2.5,an important indicator of air quality.The experiments show that using the evaluation metrics RMSE and MAE,the two models outperform traditional regression methods by more than 74% and 42%,and existing deep learning methods by more than 2.8% and 14%,respectively,in the Fuzhou dataset,and outperform traditional regression methods by more than 68% and 56%,and existing deep learning methods by more than 9.7% and 5.0%,respectively,in the Shunyi District dataset in Beijing.