Leveraging pretrained transformers for enhanced air quality index prediction model

Abstract

Air pollution mitigation is essential to ensure sustainable development, as it directly affects climate change, economic productivity, and social well-being. Despite the availability of numerous prediction techniques, machine learning (ML) remains the optimal solution for forecasting air pollution. Constructing a prediction model for a region with limited data poses a challenge. This study presents a novel technique that combines temporal fusion transformer (TFT) with transfer learning to create an inventive air quality index (AQI) prediction model, effectively utilizing temporal insights and prior knowledge. The TFT is an advanced deep neural architecture engineered to enhance time series forecasting through the fusion of sequence modelling and global temporal patterns. By fusing TFT with transfer learning, the research pioneers a fresh approach to AQI prediction for region with data scarcity issue, capitalizing on cross-domain knowledge transfer. Utilizing meteorological and pollutant data from the Cochin region, a hybrid AQI prediction model is constructed through TFT and transfer learning. Employing a preexisting TFT model trained on Trivandrum data, transfer learning technique is utilized to adapt the model for predicting AQI in the Cochin region. The study demonstrates that integrating TFT with transfer learning yields superior accuracy compared to an exclusive TFT-based approach.