A Multimodal Approach to Improve the Prediction Capabilities of Deep Learning Models for Multivariate Time Series—A Case of Subtropical Urban Air Quality

Abstract

Environmental engineering plays a critical role in managing air quality services, which are of daily concern to the public, particularly as climate change alters the factors affecting air quality. Within this context, our study introduces a comprehensive approach that emphasizes predictive models relying on multivariate time-series data. By integrating data from various sources and modalities, we propose a multimodal deep learning method to enhance traditional unimodal models. This study includes a review of existing literature, the preparation of relevant datasets, the development of robust models, and extensive evaluations. The experiments feature a case study focused on air quality services in a subtropical city, aiming to provide insights for improving prediction models. The integrated multimodal approach offers a better understanding of environmental conditions by combining data from automatic air quality monitors, meteorological stations, the European Centre for Medium-Range Weather Forecasts reanalysis data, as well as public welfare information and societal disruption reports. The analysis also considers weather-related alerts, such as typhoon and rainstorm warnings, which lead to school closures and city-wide suspensions. The model incorporates emission sources and upwind areas. Preliminary causality tests confirm that augmented feature space to encompass upstream areas enhances the model analytical capability. Downstream pollution and environmental conditions are significantly influenced by socio-economic activities in upwind areas. Granger causality and Diebold-Mariano tests highlight the importance of public welfare information and societal disruption reports, addressing a critical gap in this field.