Prediction of PM2.5 Concentration Based on CNNLSTM Deep Learning Model

Abstract

Rational prediction of $PM_{2.5}$ concentration can effectively prevent and control atmospheric environmental pollution. To improve the accuracy of short-term $PM_{2.5}$ concentration prediction, the paper proposes a combined CNN-LSTM prediction model combining CNN and LSTM networks. The model first automatically extracts the spatial features of the dataset set using a CNN and a one-dimensional convolutional kernel function, and then uses a multilayer LSTM network to capture the time-dependent features of the sequence, then introduces a Dropout layer and trains the model with the Adam optimization algorithm mechanism to improve the operational efficiency. Finally, a deep neural network with a single hidden layer is used in the fully connected layer to fit and predict the data and output the predicted value. The paper predicts $PM_{2.5}$ concentrations using Beijing air pollutant concentration data and historical meteorological data from 2014-1-1 to 2022-7-5 to fully extract the spatial and temporal characteristics of multivariate nonlinear series. The results show that the optimization of the CNN-LSTM model on the LSTM model can provide a more accurate data basis, which is used in formulating air pollution prevention and control countermeasures.