Evolving Compact Prediction Model for PM2.5 level of Chiang Mai Using Multiobjective Multigene Symbolic Regression

Abstract

In recent years, fine particulate matter (PM2.5) has caused economic and health-related adversities to people of Northern Thailand. An accurate predictive model would allow residents to take precautions for their safeties. Also, a human-readable predictive model can lead to better understandings of the issues. In this paper, we use multigene symbolic regression, a genetic programming (GP) approach, to create predictive models for PM2.5 levels in the next 3 hours. This approach creates mathematical models consists of multiple simpler trees for equivalent expressiveness to conventional GP. We also used Non-dominated Sorting Genetic Algorithm-II (NSGA-II), a multiobjective optimization technique, to ensure accurate yet compact models. Using pollutants and meteorological data from Yupparaj Wittayalai monitoring station, combined with satellite-based fire hotspots data from Fire Information of Resource Management System (FIRMS), our approach has created compact human-readable models with better or comparable accuracies to benchmark approaches, as well as identifies possible nonlinear relationships in the dataset.