Recurrent neural network for multi-steps ahead prediction of pm10 concentration

Abstract EN

This paper describes the development of a Multi Layer Perceptron (MLP) recurrent neural network based model to perform a multi-steps ahead prediction of pollutant concentration.

Receiving the latest k measurements of pollutant concentration and the meteorological parameters the model is able to predict the next k hourly concentration.

The model has been applied to predict the pmio (Particulate Matter with an aerodynamic diameter ofio micrometer) concentration in Annaba city, northeast of Algeria (north of Africa).

Compared with a single step prediction MIA model the recurrent model perform best for short-term prediction and give interesting performances for long-term prediction.