A deep learning model incorporating frequency domain information for ultra multi-step air pollutant forecasting: A case study of Shanghai

Abstract

Multi-step forecasting of air pollutants extends the horizon for individuals and authorities to take informed actions for mitigating potential risks. Due to the instability of air pollutants, current research primarily focuses on relatively short-term forecasting, with achieving ultra multi-step forecasting presenting a significant challenge. In response to this issue, this study proposes a novel model: Frequency Enhanced Decomposed Temporal Convolution Networks (Fed-TCN) to achieve ultra multi-step forecasting. This study applies time-frequency transformation to explore the frequency characteristics of air pollutants and extract long-term patterns. These patterns are then fed into TCN to enhance the accuracy of ultra multi-step forecasting. Extensive experiments were conducted on eight air pollutants at four monitoring stations in Shanghai. The results indicate variations in forecastable ranges for different pollutants. NO and NO_x can be forecasted up to one week, while NO₂, CO, SO₂, and O₃ require forecasting within 1–3 days (approximately 24–72 steps ahead). Furthermore, PM_2.5 and PM₁₀ can only be forecasted for short-term periods, not exceeding 12 h. When compared to baseline models, Fed-TCN achieves a 4.3%–11% lower Mean Absolute Error. Moreover, Fed-TCN provides insights into the contribution of pollutant composition patterns to forecasting accuracy. In general, daily patterns, semi-commuting patterns, and residuals contribute 68.8%–81.7% to ultra multi-step forecasting. The proposed method is applicable for ultra multi-step forecasts of other regions and different types of air pollutants.