Prediction of PM2.5 concentration based on the weighted RF-LSTM model

Accurate prediction of PM2.5 concentrations can provide a solid foundation for preventing and controlling air pollution. When the Long Short-Term Memory (LSTM) is applied to predict PM2.5 concentration, the influential factors strongly correlated with PM2.5 concentration are directly fed into the LSTM network. However, the influence of these factors on PM2.5 concentration is different. To address this issue, a weighted Random Forest (RF)-LSTM model was proposed to predict PM2.5 concentration for the next six hours in this study. This model first uses the RF to select the factors that are more important for predicting PM2.5 concentration and then uses a fully connected neural network to learn the weight value of each factor. Finally, the weighted data is fed into the LSTM network. The model is trained, validated, and tested using hourly air pollutant and meteorological data collected from four monitoring stations in Beijing, China, from November 1, 2019 to February 28, 2022. The prediction performance of the weighted RF-LSTM model was compared to the RF-LSTM and LSTM models. The results show that the RMSE and MAE of the weighted RF-LSTM model are the smallest, and the R2 is the largest for the next six hours' prediction of PM2.5 concentration at four stations. Compared to the LSTM model, the weighted RF-LSTM model decreases RMSE by 2.3%-5.3%, MAE by 5.6%-9.6%, and improves R2 by 2.0%-4.8%, showing that the weighted RF-LSTM model proposed in this study can achieve better prediction performance and has strong generalization ability.