A Neural Network Approach with Iterative Strategy for Long-term PM2.5 Forecasting

Exposure to a high concentration level of particulate matter (PM2.5) can increase health risks and build an unhealthy environment. It is essential to build an accurate PM2.5 forecasting model, which has an immense role in data-driven environmental decision-making strategies. Long-term forecasting of PM2.5 is a fundamental approach to early warning and instantaneous response. However, the existing long-term forecasting approach still needs further improvement. This paper proposed Multi-step Temporal Convolutional Network (MS-TCN) with the recursive strategy to improve air pollution forecasting performance. The model is developed using a multivariate PM2.5 dataset, which considers correlated variables instead of just performing univariate forecasting. The proposed multi-step model accumulates each step predicted values to conduct long-term forecasting. This forecasting model is more useful in early warning applications, which can quickly alert what will occur after multi-step conditions, which is not possible by existing single-step forecasting models. Error metrics evaluate the MS-TCN model’s performance with the pollution dataset of Beijing, China, for 46 weeks multistep ahead prediction. Quantitative comparative analysis proves the proposed MS-TCN model outperforms the baseline machine learning and deep learning models. This work has significant importance on air pollution control and prevention.