Enhancing air pollution prediction: A neural transfer learning approach across different air pollutants

Abstract

Air pollution stands out as one of the most alarming environmental challenges. It poses significant risks to human health and the environment. Accurate forecasting of air pollutant concentration levels is crucial for effective air quality management and timely implementation of mitigation strategies. In this paper, the transfer learning technique is investigated using the artificial neural network (ANN), also called multi-layer perception (MLP), to transfer knowledge across different air pollutants forecasting, and therefore, to generalize over a large set of air pollutants in the same air monitoring station. By leveraging the knowledge learned from a source forecasting task, transfer learning allows us to reduce the data requirements, speed up the training of the models, and enhance the predictive performance for different air pollutants for the target forecasting task. We present a comprehensive analysis of the transfer learning across different air pollutants in the same air monitoring station on a large dataset of air quality measurements. Our results demonstrate that transfer learning significantly improves forecasting accuracy with fewer fine-tuning data, particularly when limited labeled data is available for the target task. The findings of this study contribute to the advancement of air pollution forecasting methodologies, facilitating better decision-making processes and proactive air quality management. • Overcoming the challenge of training multiple models to forecast air pollutants. • Generalizing models across different air pollutants at the same station. • Developing a pre-trained model for P M 2 . 5 and fine-tuning it for other pollutants. • Using MLP/ANN architecture for time-series forecasting and transfer learning.

Topics & Concepts