Transfer Learning-Driven Hourly PM2.5 Prediction Based on a Modified Hybrid Deep Learning

Haze is a major problem in China’s air pollution, which not only hinders economic development, but also causes harm to people’s health. PM2.5 (fine particulate matter) is the primary cause of haze. Therefore, the timely prevention and control of haze benefit the precise forecast of PM2.5 concentration. Air quality has high-dimensional, non-linear and complex characteristics. In this paper, a modified hybrid deep learning model is proposed under the framework of transfer learning, which can solve the problem of air quality prediction in the case of sparse data. The research focuses on solving the problem of inadequate feature extraction in existing studies, and and predicts PM2.5 concentration at multiple sites. In the domain of adaptive extraction of air pollutant characteristics, long and short-term neural networks and multi-layer perceptron are used to realize the long-term dependence and nonlinear transformation of features, respectively. The learned features can be shared by PM2.5 prediction tasks at multiple sites. The channel and spatial attention mechanisms are added to extract the key information in the representation target. In the whole network, the residual neural unit is used to increase the depth of the network and improve prediction accuracy. This paper discusses the experimental results in Beijing dataset from 2013 to 2017 and Hengshui dataset from 2020 to 2022. Based on the findings, it shows that compared with the classical deep learning models, hybrid deep learning models and the most recent transfer learning approaches, the network can obtain higher accuracy and better robustness, especially for the prediction of sites with sparse data. The RMSE value of TL-Modified compared with TL-LSTM and TL-CNN-LSTM models decreased by 38%, 16.5% and 25.6% at different sites, respectively.