A Hybrid CNN-LSTM Model for Forecasting Particulate Matter (PM2.5)

TLDR

PM2.5 pollution poses a serious health risk, and forecasting its surface concentration is crucial for protecting public health. This study develops a hybrid CNN‑LSTM model to predict Beijing’s 24‑hour PM2.5 concentration by combining CNN feature extraction with LSTM temporal modeling. Using seven days of historical air‑quality data as input, the authors compare four models—univariate LSTM, multivariate LSTM, univariate CNN‑LSTM, and multivariate CNN‑LSTM—to forecast the next day’s PM2.5. Evaluation with MAE and RMSE shows the multivariate CNN‑LSTM achieves the lowest errors and fastest training, outperforming the other models.

Abstract

PM2.5 is one of the most important pollutants related to air quality, and the increase of its concentration will aggravate the threat to people's health. Therefore, the prediction of surface PM2.5 concentration is of great significance to human health protection. In this study, A hybrid CNN-LSTM model is developed by combining the convolutional neural network (CNN) with the long short-term memory (LSTM) neural network for forecasting the next 24h PM2.5 concentration in Beijing, which makes full use of their advantages that CNN can effectively extract the features related to air quality and the LSTM can reflect the long term historical process of input time series data. The air quality data of the last 7days and the PM2.5 concentration of the next day are first set as the input and output of the model due to the periodicity, respectively. Subsequently four models namely univariate LSTM model, multivariate LSTM model, univariate CNN-LSTM model and multivariate CNN-LSTM model, are established for PM2.5 concentration prediction. Finally, mean absolute error (MAE) and root mean square error (RMSE) are employed to evaluate the performance of these models and results show that the proposed multivariate CNN-LSTM model performs the best results due to low error and short training time.