Air quality forecasting models which are useful for atmospheric scientist and government can be built based on numerical theory or statistic theory.

The first approach outline the basic physics & dynamic process, chemical reaction of true atmosphere and the various kind of emission. The second approach use statistical tools to capture the pattern of historical meteorology, air quality dataset and train a predictive model. With evaluating the model for unseen data instances predicting, it can be used for air quality forecasting.

Many researchers choose numerical models to achieve their target as computation results based on the emulation of real world, could reproduce spatial-temporal variation of pollutants to some extent.

Machine learning is becoming the cutting-edge of data science even in all scientific field and showing its great ability in non-linear problems(secondary pollutants(O3, SOA, etc) are formed through non-linear chemistry reactions).

What are the possibilities to use a machine learning approach to improve air quality forecasting?

  • 6
    $\begingroup$ Interesting question. One of the most important parts of building a successful machine learning model is the quality and quantity of the data. You can gather a clever set of features such as longitude, latitude, day of year, season, time of day, wind speed, temperature, etc. But in the end you will be using historical data - which is very limited. At the same time, the earth system is extremely complex and responds abruptly to changes within the system which are highly non-linear. Therefore I would be very cautious when applying ML to any kind of forecasting that has to do with air quality. $\endgroup$ Commented Aug 12, 2016 at 19:12
  • 1
    $\begingroup$ Due to more than 10-15 input parameters being present , overfitting would be an issue . Care must be taken in distinguishing overfitting with good prediction . $\endgroup$
    – shrey
    Commented Dec 29, 2016 at 5:54

1 Answer 1


While the prospects are more of an economic question, I can answer that there is at least one study using machine learning methods (https://etda.libraries.psu.edu/catalog/13561nvb5011).

There are problems such as the attribution of error to physical processes are not feasible, the spatial extent of the model is very limited, the accuracy of the model is dependent on the history and integrity of the data, etc. Perhaps a hybrid, some sort of Model Output Statistics could be formed, but this would be doubly expensive.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.