Daily pollution forecast using optimal meteorological data at synoptic and local scales

TL;DR

This paper develops a framework for selecting key meteorological variables to improve PM10 pollution forecasting in Lisbon, comparing linear and non-linear models, and analyzing the impact of weather circulation patterns.

Contribution

It introduces a variable pre-selection method using stepwise regression and evaluates the effectiveness of linear and neural network models for pollution prediction.

Findings

Non-linear models perform similarly or worse than linear models.

Variable importance varies with measurement location.

Circulation Weather Types influence pollutant concentrations.

Abstract

We present a simple framework to easily pre-select the most essential data for accurately forecasting the concentration of the pollutant PM$_{10}$, based on pollutants observations for the years 2002 until 2006 in the metropolitan region of Lisbon, Portugal. Starting from a broad panoply of different data sets collected at several meteorological stations, we apply a forward stepwise regression procedure that enables us not only to identify the most important variables for forecasting the pollutant but also to rank them in order of importance. We argue the importance of this variable ranking, showing that the ranking is very sensitive to the urban spot where measurements are taken. Having this pre-selection, we then present the potential of linear and non-linear neural network models when applied to the concentration of pollutant PM$_{10}$. Similarly to previous studies for other pollutants, our validation results show that non-linear models in average perform as well or worse as linear models for PM$_{10}$. Finally, we also address the influence of Circulation Weather Types, characterizing synoptic scale circulation patterns and the concentration of pollutants.