Investigating Efficacy of RNN and its Variants for AQI Forecasting

Because of human-made activities like urbanization, construction, and industrialization, developing countries are particularly concerned about air pollution. Human health is significantly impacted by the air humans breathe. The air quality index is a tool to gauge air pollution. Previously, statistical tools were used to forecast Air Quality Index (AQI), but these tools had inherent limitations. With advancements in deep learning techniques, many variants of recurrent neural networks are being used for AQI forecasting. In this study, various models such as Recurrent Neural Network (RNN), Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and their bidirectional variations are investigated for two sites in Mumbai. The feature vector utilized as the dataset has a total of 15 variables that represent various air pollutants and meteorological parameters. A total of 5 forecasting models were trained and tested using a deep learning approach. All five models were compared using two performance metrics, Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), and it was discovered that bidirectional GRU outperformed at both stations. Due to the substantial variance in AQI levels at the Kurla station, even advanced deep models were unable to capture the pattern, resulting in lower prediction accuracy. Large amounts of historical data combined with advanced processing capabilities can produce better outcomes. Using a hybrid model or an ensemble of weak models, can too produce considerably better results.