Attention-based parallel networks (APNet) for PM2.5 spatiotemporal prediction

Urban particulate matter forecast is an important part of air pollution early warning and control management, especially the forecast of fine particulate matter (PM2.5). However, the existing PM2.5 concentration prediction methods cannot effectively capture the complex nonlinearity of PM2.5 concentration, and most of them cannot accurately simulate the temporal and spatial dependence of PM2.5 concentration at the same time. In this paper, we propose an attention-based parallel network (APNet), which can extract short-term and long-term temporal features simultaneously based on the attention-based CNN-LSTM multilayer structure to predict PM2.5 concentration in the next 72 h. Firstly, the Maximum Information Coefficient (MIC) is designed for spatiotemporal correlation analysis, fully considering the linearity, non-linearity and non-functionality between the data of each monitoring station. The potential inherent features of the input data are effectively extracted through the convolutional neural network (CNN). Then, an optimized long short-term memroy (LSTM) network captures the short-term mutations of the time series. An attention mechanism is further designed for the proposed model, which automatically assigns different weights to different feature states at different time stages to distinguish their importance, and can achieve precise temporal and spatial interpretability. In order to further explore the long-term time features, we propose a Bi-isnvi parallel module to extract the periodic characteristics of PM2.5 concentration from both previous and posterior directions. Experimental results based on a real-world dataset indicates that the proposed model outperforms other existing state-of-the-art methods. Moreover, evaluations of recall (0.790), precision (0.848) (threshold: 151 mu g/m(3)) for 72 h prediction also verify the feasibility degrees four proposed model. The methodology can be used for predicting other multivariate time series data in the future. (C) 2021 Elsevier B.V. All rights reserved.