A hybrid CLSTM-GPR model for forecasting particulate matter (PM2.5)

PM2.5 concentration is closely related to air pollution and human health, which should be predicted accurately and reliably. In this study, we proposed a hybrid model combining convolution neural network (CNN), long-short term memory network (LSTM), and gaussian process regression (GPR), called CLSTM-GPR model to fully extract the spatial-temporal information from the PM2.5 data series to achieve precise point prediction and dependable interval prediction. To demonstrate the model's quality and dependability, the CLSTM-GPR model was applied to PM2.5 concentration prediction at two monitoring stations, and comparisons were made with CNN-GPR, LSTM-GPR, and GPR models at the same time to evaluate the point prediction accuracy and interval prediction applicability. The CLSTM-GPR model presented the best overall prediction results with R increasing by over 4.38%, R2 increasing by over 8.96%, MAE decreasing by over 5.14%, RMSE decreasing by over 4.68%, and MC decreasing by more than 17.28% compared to other models. The results show that the CLSTM-GPR model is able to produce highly accurate point predictions and appropriate prediction intervals for PM2.5 concentration prediction. Thus, the CLSTM-GPR model has great potential for predicting PM2.5 concentrations. Also, this is the first application of the CLSTM-GPR model for PM2.5 concentration prediction. Overall, this study highlights the potential of the proposed model and demonstrates its further application in PM2.5 concentration prediction.