A machine learning model integrating spatiotemporal attention and residual learning for predicting periodic air pollutant concentrations

Abstract

The variations in pollutant concentrations during tunnel construction, in urban atmosphere, and along pedestrian paths exhibit distinct periodicity. Accurate prediction of pollutant concentrations is crucial for improving the quality of the construction and living environments. Using tunnel construction scenarios as a case, this study proposes a Convolutional Neural Network and Bidirectional Long Short-Term Memory model (CNN-BiLSTM) integrating spatiotemporal attention mechanisms and residual learning. The model employs experimental and field measurement data, with Pearson correlation analysis used for preliminary data screening. The proposed model was evaluated using eight specific metrics and compared against eight baseline models. The model exhibited strong predictive performance, with R² of 0.9826 for CO and 0.9844 for PM_2.5. For 15-step CO predictions, R² was 0.9584 with MSE of 0.031. Urban-scale predictions showed R² of 0.9599 for CO and 0.9774 for PM_2.5, while traffic-related predictions were 0.9316 for CO and 0.9525 for PM_2.5, indicating improved accuracy and applicability.