Deep Learning Theory And Verification Of Time Series Air Pollution Spatial Interpolation Problem

Deep learning,as an important branch of machine learning,has been stalled several times since it was proposed due to the limitations of linear problem and gradient disappearance.In recent years,with the effective solution of problems such as gradient disappearance,deep learning has been widely studied and applied in academia and industry fields,such as data mining,air pollutant assessment and time series prediction.Kriging interpolation method has been widely used in regional air quality assessment.However,it was originally developed for geological statistics,and the definition of its core semi-variance function is based on the first law of geography,which cannot represent the physical laws of atmospheric pollutant transport,and the historical temporal dependence of the interpolation index and related factors is not considered.Therefore,the original kriging method is not effective when applied to spatial interpolation of regional air pollutants,especially in the estimation of the distribution of air pollutants in complex environments.In this paper,a series of researches are carried out on the deep learning theory of the spatial interpolation problem of time series air pollution.By using the deep learning model and improving the kernel model of kriging interpolation,the kriging interpolation technology is optimized to make it more suitable for the study of the spatial distribution of time series atmospheric pollutants.The specific work is as follows:(1)Aiming at the problem of poor spatial interpolation of regional air pollution in time series.Firstly,the diffusion principle and transmission law of regional air pollution were analyzed.Then,the research target pollutants were selected,and kriging interpolation was performed on the target pollutants of 61 stations in the study area in turn to obtain the orginal Kriging interpolation result sequence.Finally,the error sequence between the observation value and the interpolation result was obtained.(2)Aiming at the problem that the traditional kriging interpolation optimization method is not effective.Firstly,the reason is that only the kriging interpolation results are optimized but the core algorithm is not improved,and the relevant factors of target pollutants are not considered comprehensively.Then,the correlation of the target pollutant with meteorological factors and other pollutants is analyzed,and the relevant factors were selected along with the data of coordinates,elevation and time as input to train a Long ShortTerm Memory(LSTM)model to simulate gaussian diffusion model.Finally,the LSTMKriging model was obtained by replacing the semi variance function fitting method in the kriging interpolation process with LSTM,which can represent the physical mechanism of atmospheric diffusion.(3)The validity of the LSTM-Kriging model was verified by using the time-series meteorological and pollutant data of 61 air monitoring stations in Guilin,China.The results show that the Root Mean Square Error(RMSE)of LSTM-Kriging model for the interpolation of PM2.5 concentration in Guilin is 13.4202,which is 22.27%lower than the traditional optimization method.Compared with the optimization model based on Autoregressive Integrated Moving Average model(ARIMA),Land Use Regression model(LUR),Gated Recurrent Unit(GRU),Attention-based Parallel Network(APNet)and Attention mechanism-based LSTM Densely Connected Convolution Network(DCCNALSTM),the Mean Absolute Error(MAE)and Symmetric Mean Absolute Percentage Error(SMAPE)of the LSTM-Kriging model are 12.1988 and 0.1927%,respectively.The interpolation accuracy is better than the reference geostatistical optimization model,machine learning optimization model and other simple deep learning optimization models.In this paper,a Kriging optimization method for spatial interpolation of time series air pollution was proposed.In this method,the semi-variance function was optimized by using the time-series deep learning theory and the Gaussian diffusion theory of atmospheric pollution,and the Kriging interpolation structure was preserved completely.Experimental results show that the method proposed in this paper performs well in the spatial interpolation of time series air pollution.This improved method of semivariance function can provid reference for the applicability research of Kriging in other fields.