Abstract
The swift decline in the air quality in South Asian mega cities, especially Delhi, presents significant threats to human health owing to elevated concentrations of particulate matter (PM2.5) resulting from dense populations, heavy traffic, and industrial emissions. Precise and efficient prediction of air quality is essential for successful mitigation and policy formulation. This research introduces an innovative predictive framework, AquaWave-BiLSTM, that combines sophisticated feature extraction and optimization methods to enhance multi-station air quality forecasting in Delhi. Hourly air quality and meteorological data were gathered from six monitoring sites from June 2018 to October 2019. The proposed model integrates Wavelet Transform for frequency pattern extraction, Principal Component Analysis (PCA) for dimensionality reduction, and A hybrid Aquila Optimizer and Arithmetic Optimization (AOAOA) for the selection of pertinent features. A Bidirectional Long Short-Term Memory (Bi-LSTM) network is utilized to simulate temporal interdependence. The AquaWave-BiLSTM framework demonstrated exceptional predictive accuracy, with a Mean Squared Error (MSE) of 0.00065, a Mean Absolute Error (MAE) of 0.04566, a Root Mean Square Error (RMSE) of 0.02523, and an R² value of 0.9494, surpassing conventional methodologies. Furthermore, the model exhibited computational efficiency with an average execution time of 20.57 seconds. The Wilcoxon Signed-Rank Test statistically validated the relevance of the suggested feature extraction and selection method for all monitoring stations. The AquaWave-BiLSTM framework enables efficient, interpretable air quality forecasting, with SHAP clarifying feature contributions.