Abstract
In areas that are urbanizing quickly, air pollution continues to be a major problem because it can lead to enormous dangers to human health and also to environmental sustainability. This paper suggests two strong deep learning models for precise prediction of Air Quality Index (AQI) that makes use of current atmospheric as well as meteorological data so as to raise public awareness, guide policy and also to encourage proactive citizen action. Hybrid Long Short-Term Memory–Gated Recurrent Unit (LSTM–GRU) architecture and a residual bidirectional LSTM (BiLSTM) network with 1×1 convolutional skip connections for enhanced gradient flow and multi-scale representation are the two refined deep learning architectures that are created. For capturing autocorrelations and also short-term dependencies in AQI dynamics both models make use of lagged pollutant data in conjunction along with 27 engineered features, that includes rolling and exponential statistics, indices of heat-humidity, metrics related to rate-of-change, ratios of pollutant concentration as well as cyclical encodings of temporal variables (such as day of week, hour of day). To make sure about steady convergence and robust generalization, the models were trained using the Adam optimizer with Huber loss, adaptive learning-rate decay and early stopping. Inputs are quantile-normalized. By integrating physics- based feature engineering and relevant domain knowledge, the method guarantees organized transparency in the modelling procedure. It also successfully records changes related to time, reduces data interference and adjusts to the changing characteristics of city settings, leading to more trustworthy and understandable forecasts. However, this also helps in creation of intelligent as well as scalable forecasting tools for prediction of AQI for public health dashboards, early warning systems and also for infrastructures of smart city.
Keywords: Air Quality Index, Conv1D, Deep Learning, Forecasting, LSTM–GRU, Residual BiLSTM.