Accurate and interpretable PM2.5 prediction based on GC-AE-RegLSTM

Air pollution seriously affects public health and socially sustainable development. It is imperative to accurately predict PM2.5 concentrations to alert the public and take preventive measures in advance. However, current researches on air quality prediction often ignored model complexity and interpretability issues. To solve the above problems, this paper proposes a framework based on Granger Causality Test (GC), Autoencoder (AE), and Long Short-Term Memory network (LSTM) with L2 regularization, called GC-AE-RegLSTM model. Firstly, to enhance model interpretability, the Granger causality test is applied to analyze feature relationships and determine the optimal LSTM input time series length. Secondly, to simplify the data structure and extract potential features, AE is utilized for feature learning and dimensionality reduction. Then, the LSTM model for time series prediction is built from the encoded features to better capture the dynamic characteristics of PM2.5 concentration changes. To further improve the robustness of the overall model and prevent overfitting, L2 regularization is introduced to each LSTM layer. Finally, the proposed model is trained and validated by the actual weather data in Beijing, and the model parameters are optimized by orthogonal experimental design. The experimental results show that the GC-AE-RegLSTM model performs well in predicting PM2.5 concentrations, which is significantly better than other common prediction models, with its evaluation indicators (R2, RMSE, MAE, MRE) optimized by at least 2.08%, 2.17%, 5.30%, and 7.09%, respectively. In addition, the Friedman test and Nemenyi test show that the GC-AE-RegLSTM model gets relatively better prediction accuracy than other models.