Multi-view Stacked CNN-BiLSTM (MvS CNN-BiLSTM) for urban PM2.5 concentration prediction of India's polluted cities

The existence of PM2.5 poses a substantial threat to both human well-being and ecosystems. The quantification of PM2.5 is a pressing global issue. These little particles can swiftly enter the respiratory system and penetrate the lungs profoundly, resulting in a range of health problems such as respiratory disorders, cardiovascular diseases, and early mortality. Research indicates hybrid deep learning (DL) models outperform individual DL models such as CNN, RNN, GRU, LSTM, and BiLSTM in predicting the PM2.5 pollutant in time series data. However, these hybrid models still need to attain effective performance. This study has proposed a hybrid stacked CNN Bi-LSTM model architecture that incorporates many data perspectives related explicitly to seasonal repeats to generate numerous models. This approach is called Multi-view Stacked CNN Bidirectional-LSTM (MvS CNN-BiLSTM). The suggested approach has been deployed on seventeen individual time series data sets of PM2.5 levels in highly polluted cities in India, using stand-alone deep learning models. The effectiveness of the proposed approach has been evaluated by comparing its performance using the RMSE and MAPE metrics. The suggested model has obtained an average enhancement compared to stand-alone DL models on all datasets as follows: RMSE: 7.11% (CNN), 5.08% (RNN), 3.80% (GRU), 5.57% (LSTM), and 4.05% (BiLSTM). Additionally, the MAPE values are 27.16% (CNN), 28.52% (RNN), 26.22% (GRU), 27.22% (LSTM), and 23.11% (BiLSTM). In addition, a non-parametric statistical analysis (Friedman and Holm's) was conducted, demonstrating that the proposed MvS CNN-BiLSTM outperforms both performance metrics significantly. The proposed approach has the potential to improve policy, human health, and ecosystems by focusing on actions that prioritize the reduction of pollutant emissions and the enhancement of air quality management.