Predicting PM2.5 concentration with enhanced state–trend awareness and uncertainty analysis using bagging and LSTM neural networks

Abstract

Monitoring air pollutants, particularly PM2.5, which refers to fine particulate matter with a diameter of 2.5 µm or smaller, has become a focal point of environmental protection efforts worldwide. This study introduces the concept of state–trend awareness, which is widely employed in big data analytics to enhance global threat identification, understanding, and response capabilities. We applied this approach to the prediction of PM2.5, utilizing its capacity to provide holistic insights and support decisions in dynamic environments. We conducted in-depth analyses of extensive historical data to forecast the future concentration trends. By combining a long short-term memory (LSTM) neural network with a bagging ensemble learning algorithm, our developed model demonstrated superior accuracy and generalization compared to those of traditional LSTM and support vector machine (SVM) methods, reducing errors relative to SVM-LSTM by 12%. We further introduced interval prediction to address forecasting uncertainties, not only providing a specific PM2.5 but also forecasting the probability ranges of its variations. The simulation results illustrate the effectiveness of our approach in improving the prediction accuracy, enhancing model generalization, and reducing overfitting, thereby offering a robust analytical tool for environmental monitoring and public health decision-making.