A Platform for Air Quality Forecast
Oral Presentation
Prepared by F. Ramos1, V. Guizilini2
1 - National ICT Australia, 13 Garden Street, Eveleigh, Sydney, NSW, 2015, Australia
2 - NICTA, 13 Garden Street, Eveleigh, Sydney, NSW, 2015, Australia
Contact Information: [email protected]; 129-036-7058
ABSTRACT
This paper discusses the theoretical and implementational aspects of developing a platform for air quality control. The goal is to obtain, from a set of sparse measurements on a determined area over time, predictive values for PM 10 and PM 2.5 pollutants. This prediction occurs both in space and time, allowing the system to produce estimates for areas where there are no sensors and also in the near future. We use the Gaussian process framework to generate a starting model based on historical data, which is then continuously updated as new measurements becomes available. A novel training methodology is proposed, named structural cross-validation, that is shown to improve accuracy by minimizing an error function while maintaining the spatial-temporal structure of available data. We also explore the use of other environmental variables, such as temperature, humidity and wind, to improve the accuracy of predictions over PM10 and PM2.5. Initial tests were conducted on the Hunter Valley area, with 14 sensors distributed over an area of roughly 50×80 km providing hourly measurements. The resulting model is currently being developed into a working platform to be deployed by the Australian government, as a way to monitor air quality in populated areas.
Oral Presentation
Prepared by F. Ramos1, V. Guizilini2
1 - National ICT Australia, 13 Garden Street, Eveleigh, Sydney, NSW, 2015, Australia
2 - NICTA, 13 Garden Street, Eveleigh, Sydney, NSW, 2015, Australia
Contact Information: [email protected]; 129-036-7058
ABSTRACT
This paper discusses the theoretical and implementational aspects of developing a platform for air quality control. The goal is to obtain, from a set of sparse measurements on a determined area over time, predictive values for PM 10 and PM 2.5 pollutants. This prediction occurs both in space and time, allowing the system to produce estimates for areas where there are no sensors and also in the near future. We use the Gaussian process framework to generate a starting model based on historical data, which is then continuously updated as new measurements becomes available. A novel training methodology is proposed, named structural cross-validation, that is shown to improve accuracy by minimizing an error function while maintaining the spatial-temporal structure of available data. We also explore the use of other environmental variables, such as temperature, humidity and wind, to improve the accuracy of predictions over PM10 and PM2.5. Initial tests were conducted on the Hunter Valley area, with 14 sensors distributed over an area of roughly 50×80 km providing hourly measurements. The resulting model is currently being developed into a working platform to be deployed by the Australian government, as a way to monitor air quality in populated areas.
