Research On Time Series Forecasting Of Air Quality Index Based On Deep Learning

The concentration of air pollutants is an important reference data for measuring air quality indicators,and over time,a large amount of historical data on changes in the concentration of air pollutants can be collected in the observation station.If the air pollutant concentration is accurately predicted in time series,and the data is predicted and analyzed,the comprehensive assessment and alarm function of air quality can be realized.The air quality index data has the characteristics of more data noise and drastic changes,which greatly increases the difficulty of prediction.In addition,there are many air pollutants.Although a single time series forecast is relatively simple and efficient,some influencing factors may be ignored.The main research content of this article is to study the combination of practical applications,design time series forecasting methods according to the specific characteristics of the data,and realize effective and accurate forecasting strategies.The main research contents of this article include:(1)Aiming at the characteristics of high noise and drastic changes in air quality data,a new SWT-NLSTM based on Stationary Wavelet Transform(SWT)and Nested Long Short-Term Memory(NLSTM)was proposed.The model was used for univariate air quality indicator time series forecasting.The model uses the method of stable wavelet transform to decompose the original data into two types of highfrequency and low-frequency sub-sequences,which are respectively learned by nested long short-term memory deep learning models.Through the wavelet transform to separate the mixed data features in the original data sequence,the deep learning model can better learn the inherent laws of time series data,and the prediction accuracy is greatly improved by the prediction results.Compared with the LSTM model,the percentage absolute error(MAPE)is reduced by about 20%.(2)Based on the SWT-NLSTM model,aiming at its shortcomings in the prediction results,a time series forecast of univariate air quality indicators based on the MODSWT-NLSTM model is proposed.This model aims to improve the problem of inaccurate data peak prediction in the SWT-NLSTM model.After analyzing the characteristics of wavelet transform data decomposition,it is speculated that the problem is caused by the failure to decompose the high-frequency sub-sequences in the original data.Aiming at the problem of insufficient decomposition of highfrequency sub-sequences by wavelet transform,the wavelet transform is improved,and the secondary decomposition of high-frequency sub-sequences is added,so as to achieve more accurate prediction of high-frequency sub-sequences in air quality data.The accuracy is further improved,and the MAPE is reduced by about 20% compared to the prediction results of the LSTM model,while reducing the number of outlier predictions.(3)In order to comprehensively analyze the correlation of various air pollutant components and realize the most ideal multivariable air quality index prediction,a framework based on Multi-Task & Multi-Channel NLSTM(MTMC-NLSTM)is proposed.)Time series forecast of multivariate air quality indicators.The model can comprehensively consider the interaction between multiple factors through the design of its multi-task framework and feature fusion layer and realize multi-variable prediction of air quality data,and process the data sub-sequences obtained by wavelet transform decomposition through multi-channel module design,thereby improving the prediction accuracy of each air pollutant composition data.This method realizes the parallel time series prediction of 6 main air pollutant component variables including PM2.5,PM10,etc.,and significantly improves the prediction accuracy,and the error is only about 10% of the LSTM model.