Comprehensive neural network forecasting system for ground level ozone in multiple regions

Abstract

A comprehensive neural network daily maximum 8 hour-ozone forecasting model was developed based on five years of data (2010-2014) collected from 50 monitoring sites from the Dallas Fort Worth, Houston-Galveston-Brazoria, Los Angeles, San Joaquin and San Diego regions. This work represents the first neural network developed to forecast ozone in multiple regions, as well as multiple sites in the same region. Previous studies have developed separate neural network models to forecast ozone at each location. Two stages of feature selection were applied to reduce input vector dimension and redundancy. These are Piecewise Linear Orthonormal Floating Search (PLOFS), and Karhunen - Loève Transform (KLT). Two possible approaches for organizing the data were tried. These are a tall file approach and a median file approach. Results showed better performance of the tall file approach. The Multilayer Perceptron (MLP) neural network used in this study showed better prediction performance compared to other existing MLP neural network approaches.