A hybrid PM2.5 interval concentration prediction framework based on multi-factor interval decomposition reconstruction strategy and attention mechanism

Abstract

With the acceleration of urban modernization, the temporal variability in air pollution has become increasingly significant. Predicting average daily pollutant concentrations no longer suffices for decision-making in public health and environmental management. Therefore, this paper proposes the TCN-LSTM-Attn model, a hybrid PM2.5 interval concentration prediction model based on decomposition reconstruction and attention mechanism. Firstly, the interval grey incidence analysis (IGIA) and the bivariate ensemble empirical mode decomposition (BEEMD) algorithm were respectively used for feature selection affecting PM2.5 concentration and decomposition of input variables. Subsequently, considering the uneven distribution of components in various influencing factors and the high computational complexity of the model, an interval reconstruction coefficient (RCI) and evolutionary clustering algorithm (ECA) effectively clustered these components. Finally, the proposed TCN-LSTM-Attn model generated the corresponding prediction results. An empirical study evaluated the model using air quality datasets from three environmental monitoring sites in different geographical locations in Beijing. The evaluation results demonstrated that compared to the benchmark interval prediction models, the model proposed in this paper exhibited substantial improvements across all five interval evaluation criteria, with average reductions of 15.19%, 15.30%, 32.07%, 16.63%, and 33.87%, respectively. These results highlight superior performance in terms of prediction accuracy and stability across the board. The integrated attention mechanism hybrid model proposed in this paper supports routine urban air quality warnings.