Long Sequence Prediction Of Air Pollutant Concentration Based On Deep Learning

With the rapid advancement of industrialization and urbanization,the problem of air pollution has become increasingly prominent,and frequent polluted weather has brought many negative impacts to society and huge economic losses.Therefore,in order to reduce losses,it is very important to study effective methods for forecasting the concentration of air pollutants.For the long sequence prediction of air pollutant concentration,the existing researches mainly has the following three problems: First,the existing researches pays more attention to short-term prediction and rarely discusses the methods of long-term or long sequence prediction;Second,due to the disappearance of gradient and memory constraints,most existing methods are difficult to use all the context information of time series,and the long sequence prediction performance of the model is insufficient;Finally,the existing methods are difficult to reasonably model and analyze the spatiotemporal features of pollutant data,making it difficult to fully utilize the deep features behind pollutant data.A long sequence prediction model MSformer based on multi-scale learning is proposed.Aiming at the problem of a few existing research and insufficient prediction ability,a Transformer like architecture is proposed to better extract the long-term dependence of time series features.MSformer uses a multi-scale self-attention mechanism to extract multi-time scale pollutant features to calculate context information;a multi-scale-based encoding and decoding strategy is used to combine pollutant feature information with different time granularities to generate multi-scale features to enhance the decoding process.A spatio-temporal feature extraction method for air pollutant data is proposed.Aiming at the problem of how to model the temporal and spatial characteristics of pollutant data,according to the obvious correlation of pollutant data on time and spatial scales,multiple onedimensional convolutional neural networks are designed to learn the local trend characteristics of multivariate variables,and the attention of graphs is designed.The network focuses on the spatial correlation between multi-site pollutant concentrations and fuses meteorological data to extract spatiotemporal features of multi-site pollutant data.An end-to-end model for long sequence prediction of air pollutant concentration is proposed.By combining the feature extraction method based on spatio-temporal information fusion and the long sequence prediction model based on multi-scale learning,the deep spatiotemporal feature representation behind the pollutant data is extracted and combined with the long sequence prediction model to achieve higher accuracy predictions.In this thesis,extensive experimental evaluations of the above methods are carried out on the ETTm1 and Taiwan datasets.The experimental results show that the methods proposed in this thesis have achieved quantitative improvements for the long-sequence prediction of atmospheric pollutant concentrations.