An intelligent AQI prediction method using transfer learning and an IASO-optimized DCM-DESN under multi-source data: A case study of four Chinese cities

Abstract

Accurately understanding the real-time evolution of the Air Quality Index (AQI) is essential for atmospheric pollution prevention and urban public health management. Although scholars have proposed numerous AQI prediction models, most have not fully considered AQI data's noise characteristics. Moreover, previous models mainly focus on a single region and do not effectively address cross-regional prediction issues. To tackle these challenges, this paper proposes an intelligent AQI monitoring framework: using transfer learning and an improved atomic search optimization (ASO) algorithm optimized deep echo state network with a dynamic activation mechanism (DCM-DESN) under multi-source data. Firstly, a data processing method combining complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and phase space reconstruction (PSR) is designed for the collected multi-source AQI data. CEEMDAN is applied to adaptively decompose the data, significantly reducing noise effects, followed by PSR on the denoised data to identify the dynamic characteristics of the AQI time series data. Secondly, we design a deep learning model of DCM-DESN for AQI prediction. By adjusting the nonlinear connection mode of DESN through a dynamic activation mechanism, it reduces the model's dependence on the number of reservoir neurons. Meanwhile, an improved ASO (IASO) is proposed to optimize the network weights of DCM-DESN. Finally, daily data from Xi’an, Wuhan, Changsha, and Shenzhen covering the period from January 1, 2021 to March 24, 2024 were collected to construct multiple cross-city transfer-learning tasks. Historical pollutant concentrations and related environmental variables available up to the current day were used to directly predict the AQI one day ahead. For each task, the IASO-DCM-DESN is first pretrained on the source-city data. The pretrained reservoir-based model is then transferred to the target city and frozen, while the output weights are fine-tuned using a limited target-domain adaptation set. The proposed model’s performance across various tasks is validated using MAE, RMSE, and R² metrics, demonstrating excellent performance.