ConvFormer-KDE: A Long-Term Point–Interval Prediction Framework for PM_2.5 Based on Multi-Source Spatial and Temporal Data

Abstract

Accurate long-term PM2.5 prediction is crucial for environmental management and public health. However, previous studies have mainly focused on short-term air quality point predictions, neglecting the importance of accurately predicting the long-term trends of PM2.5 and studying the uncertainty of PM2.5 concentration changes. The traditional approaches have limitations in capturing nonlinear relationships and complex dynamic patterns in time series, and they often overlook the credibility of prediction results in practical applications. Therefore, there is still much room for improvement in long-term prediction of PM2.5. This study proposes a novel long-term point and interval prediction framework for urban air quality based on multi-source spatial and temporal data, which further quantifies the uncertainty and volatility of the prediction based on the accurate PM2.5 point prediction. In this model, firstly, multi-source datasets from multiple monitoring stations are preprocessed. Subsequently, spatial clustering of stations based on POI data is performed to filter out strongly correlated stations, and feature selection is performed to eliminate redundant features. In this paper, the ConvFormer-KDE model is presented, whereby local patterns and short-term dependencies among multivariate variables are mined through a convolutional neural network (CNN), long-term dependencies among time-series data are extracted using the Transformer model, and a direct multi-output strategy is employed to realize the long-term point prediction of PM2.5 concentration. KDE is utilized to derive prediction intervals for PM2.5 concentration at confidence levels of 85%, 90%, and 95%, respectively, reflecting the uncertainty inherent in long-term trends of PM2.5. The performance of ConvFormer-KDE was compared with a list of advanced models. Experimental results showed that ConvFormer-KDE outperformed baseline models in long-term point- and interval-prediction tasks for PM2.5. The ConvFormer-KDE can provide a valuable early warning basis for future PM2.5 changes from the aspects of point and interval prediction.

Keywords

• fine particulate matter
• long-term point prediction
• interval prediction
• convolutional neural network
• transformer
• kernel density estimation