A Cyclostationary Neural Network Model for the Prediction of the NO 2 Concentration Monica Bianchini, Ernesto Di Iorio, Marco Maggini, Chiara Mocenni, Augusto Pucci Dipartimento di Ingegneria dell’Informazione Via Roma 56, 53100 Siena (ITALY)

Air Pollution Problem Nitrogen oxide (NO x = NO + NO 2 ) emissions are among the most important factors affecting the air quality in urban areas Traffic is the main problem on a local urban scale Modeling efforts to predict and control the NO x concentrations Development of tools for pollution management

Nitrogen oxide (NO x = NO + NO 2 ) emissions are among the most important factors affecting the air quality in urban areas

Traffic is the main problem on a local urban scale

Modeling efforts to predict and control the NO x concentrations

Development of tools for pollution management

Project Goals Build an efficient neural based model for the prediction of the NO 2 concentration First prediction approximation for an early warning Independence from exogenous data Modeling the NO 2 time series only based on the past concentrations of NO and NO 2

Build an efficient neural based model for the prediction of the NO 2 concentration

First prediction approximation for an early warning

Independence from exogenous data

Modeling the NO 2 time series only based on the past concentrations of NO and NO 2

The Cyclostationary Neural Network Model Correlation of past NO and current NO 2 (daily periodicity) NO 2 (t) follows a cyclostationary dynamics (period T = 24) CNN model composed by 24 MLP blocks one for each random variable of the cyclostationary process where T = 24 and f k with k = (t mod T) + 1 , represents the k–th approximation function realized by the k–th MLP block

Correlation of past NO and current NO 2 (daily periodicity)

NO 2 (t) follows a cyclostationary dynamics (period T = 24)

CNN model composed by 24 MLP blocks one for each random variable of the cyclostationary process

where T = 24 and f k with k = (t mod T) + 1 , represents the k–th approximation function realized by the k–th MLP block

Model Architecture

Experimental Setup Data gathered by ARPA Lombardia (northern Italy) ARPA supplies a real–time air pollution monitoring system composed by mobile and fixed stations Dataset made up by NO and NO 2 concentrations detected hourly by the station number 649 (Brescia–Broletto) Performance measures: mean absolute error and mean square error

Data gathered by ARPA Lombardia (northern Italy)

ARPA supplies a real–time air pollution monitoring system composed by mobile and fixed stations

Dataset made up by NO and NO 2 concentrations detected hourly by the station number 649 (Brescia–Broletto)

Performance measures: mean absolute error and mean square error

Experimental results – err 2 months

Experimental results – err 12 months

Experimental results – mse 2 months

Future Works CNN hardware implementation on NO x sensors Management of multiple data from different sensors Testing on other urban area datasets Testing on wider datasets

CNN hardware implementation on NO x sensors

Management of multiple data from different sensors

Testing on other urban area datasets

Testing on wider datasets