PM2.5 Concentration Forecasting in the Kolkata Region With Spatiotemporal Sliding Window Approaches

Abstract

Amid rapid urbanization in Kolkata, a city in India, forecasting air pollution, particularly PM2.5 concentrations, stands as a critical challenge. In response, this study introduces a novel approach harnessing a hybrid data processing component that enables lightweight spatiotemporal models, including ConvLSTM and BiConvLSTM, to demonstrate noteworthy competitiveness. This component addresses challenges related to data availability, resource-intensive training, and the integration of meteorological factors. Leveraging a relatively small yet comprehensive dataset focusing on the Kolkata region (from February 2, 2020, to April 11, 2023), this method effectively captures nuanced pollution dynamics with spatial and temporal considerations. The results of this approach show the BiConvLSTM model consistently outperformed several models, demonstrating its superiority at stations like Bidhannagar and Fort William, with R2 values 0.901 and 0.899 respectively. The ConvLSTM model, while not as dominant overall, exhibited significant advantages in specific scenarios, particularly evident in R2 values reaching up to 0.890 at Jadavpur and 0.887 at Victoria. These findings highlight the efficacy of the proposed hybrid data processing component in enabling lightweight spatiotemporal models to achieve competitive performance in forecasting PM2.5 concentrations across urbanized landscape of Kolkata.

Keywords

• Convolutional long short-term memory networks
• deep learning
• fine-grained air pollution estimation and forecast
• Kolkata
• India
• spatial-temporal data