The present study proposes a hybrid BiGRU-1DCNN model to predict PM2.5 levels in Delhi, India, by leveraging data from multiple monitoring stations. The proposed model incorporates Bidirectional Gated Recurrent Units (BiGRU) and a one-dimensional Convolutional Neural Network (1DCNN) to capture both temporal dependencies and spatial correlations in PM2.5 data. The model’s performance is evaluated through both single-station (SS) and spatio-temporal correlation (STC) approaches. Results demonstrate that the hybrid BiGRU-1DCNN model outperforms traditional deep learning models in both SS and STC scenarios. Specifically, it achieved a minimal Root Mean Square Error (RMSE) of 15.75, Mean Square Error (MSE) of 248.04, Mean Absolute Error (MAE) of 9.04, and Mean Absolute Percentage Error (MAPE) of 13.31 at the Jawaharlal Nehru Stadium (JNS) station. For comparison, the univariate SS model for the Major Dhyan Chandra National Stadium (MDCNS) station produced an RMSE of 17.31, MAE of 10.03, MAPE of 14.50, and MSE of 299.59. The non-parametric Friedman ranking further corroborated the superior performance of the hybrid BiGRU-1DCNN model, with it achieving the highest ranking across all performance metrics compared to other models. These results highlight the potential of the ST BiGRU-1DCNN model as a robust tool for air quality forecasting and public health risk mitigation in highly polluted urban environments like Delhi.
Air pollution, particularly fine particulate matter (PM2.5), has become a global environmental and public health crisis. Rapid urbanization, industrialization, and increased human activity in major cities, compounded by climate change, have significantly deteriorated air quality. Air pollution is hazardous to human civilization and one of the most significant environmental issues, attracting global attention. The dataset was chosen for its comprehensive spatial coverage of Delhi, allowing for an in-depth spatiotemporal analysis of air quality trends. The CPCB is a government agency responsible for monitoring and regulating air quality across India, and its dataset is considered one of the most reliable sources of air pollution data for the region. Delhi, the capital of India, spans a geographical area of approximately 1,484 km2, encompassing 11 districts. As one of the most polluted cities globally, it experiences extreme air quality fluctuations due to vehicular emissions, industrial activities, biomass burning, and meteorological conditions.
The BiGRU-1DCNN model is implemented using the keras and tensorflow frameworks within the jupyter notebook development environment, chosen for their versatility and ease of model development. The model architecture employs the LeakyReLU activation function with an alpha value of 0.7, enhancing the network’s ability to learn from a wider range of activations. It utilizes a GRU layer with 100 units to capture sequential dependencies and a 1d-conv layer with 128 filters to extract local temporal features. The model is trained with a batch size of 16 using the nadam optimizer, set with a learning rate of 0.01 to balance convergence speed and stability. The experiments were conducted on a microsoft windows 10 pro system equipped with an Intel core i5-9500 CPU running at 3.0 GHz, featuring six physical cores and six logical processors, and supported by 16 GB of RAM. This configuration provided a capable yet accessible computational environment for model development and evaluation.