Prediction of particulate matter 2.5 concentration based on attention mechanism and convolutional BiLSTM network

Abstract

PM2.5 is one of the most harmful pollutants in air pollution, and long-term exposure to elevated PM2.5 concentrations can cause significant health issues. Hence, accurate PM2.5 concentration prediction is critical for early warning systems and public health safeguards. Recently, deep learning methods for PM2.5 concentration prediction have garnered considerable attention and demonstrated significant advancements. However, existing models still exhibit limitations, particularly in the extraction of temporal and spatial features from multivariate PM2.5 concentration sequences. To address these challenges, we propose a Conv-attention-BiLSTM-attention model for PM2.5 concentration prediction, utilizing a Convolutional Layer, Attention Mechanism, and Bidirectional Long Short-Term Memory (BiLSTM). The model employs a convolutional layer to capture spatial features of multivariate PM2.5 sequences, while BiLSTM extracts temporal features, and the attention mechanism identifies key information within the sequence. During input feature selection for the multivariate model, the PM2.5 concentration from this site, along with data from 11 additional sites, was selected as input based on experimental results. Finally, the model’s effectiveness and superiority are validated through comparative analysis with other models.