An attention-based domain spatial-temporal meta-learning (ADST-ML) approach for PM2.5 concentration dynamics prediction

Abstract

Because PM2.5 causes serious health impacts, accurately predicting PM2.5 concentration is important for air pollution mitigation. The problem is to capture its dynamic evolution and continuously predict changes based on historical observations. However, the task of prediction is extremely challenging due to the spatial-temporal correlation of PM2.5 (i.e., dynamic changes over time and geographic location) and spatial-temporal uncertainty (i.e., it is affected by multiple factors in temporal and spatial distribution). We propose a novel approach of attention-based domain spatial-temporal meta-learning (ADST-ML), which can train the distribution of PM2.5 prediction tasks in an area through historical PM2.5 data, meteorological data, and temporal feature data, thereby learning a strategy that can quickly adapt to tasks such as predicting the unobservable process by which the concentration of PM2.5 changes. Our prediction model has the following innovations: (1) ADST-ML has a novel network architecture design, using convolutional neural networks (CNN), an attention mechanism, and a long short-term memory (LSTM) network model to adaptively extract the spatial-temporal dependence of PM2.5 and other multivariate factors; (2) ADST-ML captures the dynamic changes of PM2.5 prediction tasks through variational inference; (3) ADST-ML designs meta-training and testing algorithms based on Bayesian meta-learning to represent model uncertainty and knowledge transfer from training data. When conducting experiments on a dataset containing Beijing air quality inspection records, our ADST-ML outperforms baselines in terms of prediction accuracy. Since it is a regional forecasting model, it breaks through some of the limitations of forecasting based on the original site, and offers the possibility of accurate forecasting in urban areas.