| Low air quality is a significant environmental and public health concern, especially in urban areas where the environment is very conducive to the accumulation of air pollution and where large groups of people can be affected at once. Sofia, has a history of measurably poor air quality, which requires decisive and swift actions on part of city governance and population. The application of such actions need to be informed by evidence about their effect. In an attempt to monitor and eventually control air quality, the Bulgarian Executive Environmental Agency has positioned five monitoring stations scattered across the city, which measure different air pollutants and meteorological factors. However, station hardware and software are susceptible to faults that may result in extended periods of data unavailability, which may affect other services and analyses down the line. To illustrate this, we present a case study where one of the five stations experiences a hypothetical fault and stops functioning. A deep learning approach is proposed, which can infer complex temporal and spatial relations and outputs multiple-hour ahead air quality forecasts. The proposed model has an encoder-decoder architecture and three main components– spatial and temporal modules, supported by spatio-temporal attention mechanisms, combined with long short-term memory networks. Experimental results demonstrate that the proposed approach performs well, highlighting its ability to simulate air quality station's measurements. These promising results bring us a step closer to reliable and data-driven air quality forecasting, supporting informed decision-making for urban planning and public health initiatives. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.