Multivariate Air Quality Forecasting with Residual Nested LSTM Neural Network Based on DSWT

With the continuous deterioration of air quality and the increasingly serious environmental problem of air pollution, accurate air quality prediction is of great significance for environmental governance. Air quality index (AQI) prediction based on deep learning is currently a hot research topic. The neural network model method currently used for prediction has difficulty effectively coping with the high volatility of AQI data and capturing the complex nonlinear relationships and long-term dependencies in the data. To address these issues, this paper proposes multivariate air quality forecasting with a residual nested LSTM neural network based on the discrete stationary wavelet transform (DSWT) model. Firstly, the DSWT data-decomposition technique decomposes each AQI data point into multiple sub-signals. Then, each sub-signal is sent to the NLSTM layer for processing to capture the temporal relationships between different pollutants. The processed results are then combined, using residual connections to mitigate issues of gradient vanishing and explosion during the model training process. The inverse mean squared error method is combined with the simple weighted average method, to serve as the weight-update approach. Back propagation is then applied, to dynamically adjust the weights based on the prediction accuracy of each sample, further enhancing the model’s prediction accuracy. The experiment was conducted on the air quality index dataset of 12 observation stations in and around Beijing. The results show that the proposed model outperforms several existing models and data-processing methods in multi-task AQI prediction. There were significant improvements in mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and R square (R²).