Hybrid deep learning algorithms for forecasting air quality index using dimension reduction technique in search of precise results

Time series forecasting of multi-variant Air Quality Index data was done using sequential hybrid models. A hybrid algorithm comprised of Principal Component Analysis (PCA) merged with various deep learning approaches such as GRU (GATED RECURRENT UNIT), LSTM (LONG SHORT TERM MEMORY), and Bi-LSTM (Bi-LONG SHORT TERM MEMORY) techniques were applied for predicting the Air Quality Index (AQI) of Vishakhapatnam city. The parameters included for the valuation of the accuracy of the considered model were root mean square error (RMSE) and cosine similarity (CS).Historical dataset collected was processed through PCA containing 90%, 95%, and 99% information of entire data. This reduced the dimension of the dataset. GRU, LSTM, and Bi-LSTM techniques were applied on these data (s) carrying 90%, 95%, and 99% information of original dataset for 100 epoch(s). We observed the GRU to be the best out of the three deep learning approaches. The values of RMSE and CS were (0.0777, 0.9735), (0.0837, 0.9728) and (0.0780, 0.9740) respectively. In other case, when prediction is carried out without PCA, the result revealed that Bi-LSTM is the best of three used techniques. The RMSE and CS were found as 0.0824 and 0.9718 respectively. The experimental results expressed that the hybrid PCA-GRU model outperform in comparison with PCA-LSTM, PCA-Bi-LSTM, and standalone models like GRU, LSTM and Bi-LSTM models for 100 epoch. In order to determine the interdependency of various dependent parameter of the dataset, a loading plot was analyzed. A narrow angle was found between PM 2.5 , PM 10 , NH 3 with AQI which indicate that these parameters are positively and strongly correlated and will largely affect the value of AQI, therefore, efforts must be made to reduce these contents from the air so as to secure human life.