Air pollution is among the major threats to human well-being, highlighting the critical need for air quality monitoring, especially in urban areas. Whereas the development of lowcost pollution sensors has facilitated a widespread monitoring, a reliable anomaly detection system is required to properly characterize data for the end-users. In this paper, we propose an enhanced deep learning approach based on the A3T-GCN (Attention Temporal Graph Convolutional Network) model that accurately forecasts particulate matter PM2.5 concentrations using real past measurements from a deployed sensor network. Our proposed Enhanced-A3T-GCN embeds all the available spatial and temporal correlations within the sensor network, along with additional information regarding the sensor environment in a graph. It is shown to achieve significant performance improvement with respect to other deep learning forecasting methods, emphasizing the importance of exploiting the sensor environment-based information. Further, the achieved accurate forecasting makes it possible to detect anomalies injected at both single and multiple sensor levels.