Knowledge-Guided Temporal Representation Learning for Air Pollution Forecasting (A Case Study of Nasiriyah City)

Abstract

Air pollution prediction is a key research problem due to its direct impact on public health and the urban environment. Moreover, because air pollutants are dynamic and nonlinear, this poses a complex challenge for time-series modeling. However, while there has been significant development in the use of deep learning–based models for air quality prediction, the vast majority of existing methods are primarily data-driven, concentrating solely on reducing prediction errors in the observed value space while neglecting the modeling quality and stability of the learned temporal representations. In this work, we introduce the Knowledge-Guided Temporal Representation Learning (KG-TRL) framework for improving prediction performance. The proposed methods derive from dual-space optimization extended with temporal structure, enhancing prediction accuracy and enabling the mechanism to organize the underlying temporal representations in a dedicated space to maintain the consistency and stability of the extracted temporal components. Experimental on-air pollution data from Nasiriyah, Iraq. The results further validate that the proposed technique exhibits stable learning and significantly better predictions.