Application of Artificial Intelligence Algorithms in Urban Air Quality Monitoring and PM2.5 Forecasting

Abstract

Urban air quality monitoring systems continuously generate high-frequency and multi-variable observational data, making it difficult to conduct effective analysis solely by relying on fixed thresholds or traditional statistical models. Existing PM2.5 prediction studies still exhibit deficiencies in integrating pollutant, meteorological, temporal, and site-specific information, and do not sufficiently account for factors such as short-term fluctuations, long-term temporal dependencies, missing observations, and sensor noise. To address these limitations, this study proposes a multi-source temporal convolutional network–gated recurrent unit attention network (MTGA-Net), built upon the open-source Beijing multi-site air quality dataset. The proposed architecture combines causal time-series convolution, gated recurrent units, and a temporal attention mechanism to forecast PM2.5 levels for the next 1 hour, 6 hours, and 24 hours. Experimental results demonstrate that MTGA-Net achieves robust predictive performance across different forecasting horizons, notably outperforming the Transformer encoder in medium- and long-term prediction tasks. Further ablation experiments reveal that meteorological features, the GRU module, and the TCN module are the key factors influencing overall model performance. Importantly, the model maintains strong predictive capability even under conditions of data missingness and noise interference. In summary, MTGA-Net can effectively enhance multi-temporal air quality prediction, providing reliable support for air quality forecasting and reliability assessment under incomplete monitoring data conditions.