A Neural Network model forecasting for prediction of hourly ozone concentration in Corsica

Abstract

This paper presents the first results of a research project aimed at building a pollution peaks predictor using Artificial Neural Networks (ANNs) with data measured locally. We focus more particularly on the ozone concentration prediction in the Corsica Island at horizon “h+1”. We mainly look at the Multi-Layer Perceptron (MLP) network which is the most used of ANNs architectures both in the Environment domain and in the time series forecasting. We have demonstrated that an optimized MLP with endogenous, exogenous and time indicator inputs can forecast hourly ozone concentration with acceptable accuracy. The final results indicate that our predictor has an average Mean Absolute Percentage Error (MAPE) equal to 10.5%. Knowing that the devices measurement accuracy is around 10%, these results are considered as very convincing by “Qualitair Corse”, regional organization responsible for monitoring air quality. We have also tested in "real conditions" our predictor: indeed, several ozone pollution peaks occurred during the months of June and August 2010. While PREV'AIR, the national air quality forecasting and mapping system, cannot predict the August's peaks, it appears that our optimized MLP is able to predict them in both cases.