A novel deep learning prediction method for PM2.5 integrating improved MSTL decomposition and FATA optimization algorithm

Accurate prediction of PM2.5 time series remains challenging. To address this issue, this paper proposes a novel PM2.5 concentration prediction model named FATA-GSPMSTL-CNN-LSTM.Firstly,the optimal feature descriptor set is determined by combining Recursive Feature Elimination with Cross-Validation (RFECV) and Pearson correlation analysis. On the basis of Multiple Seasonal-Trend decomposition using Loess (MSTL), a novel adaptive decomposition algorithm—Grid Search-based Power Spectral MSTL (GSPMSTL)—is constructed by introducing spectral density analysis and grid search strategy, which is adopted to decompose the PM2.5 time series, and the dataset is reconstructed through feature data fusion. The newly developed Fata morgana optimization algorithm (FATA) is utilized to optimize the model hyperparameters for further improving prediction accuracy. Finally, the Convolutional Neural Network- Long Short-Term Memory network (CNN-LSTM) is employed to obtain the final prediction results. Considering climate, topography and seasonal factors, PM2.5 prediction and evaluation are separately conducted in heating seasons and non-heating seasons for Guangzhou and Xianyang cities. The results demonstrate that the proposed model achieves higher prediction accuracy and stability, which can provide important application value for air quality early warning and pollution control.