An Improved Weight Optimization of Hybrid Machine Learning Models for Forecasting Daily PM2.5 Concentration

Abstract

PM_2.5 is an air pollutant primarily produced by human activities, including the combustion of fossil fuels, industrial emissions, vehicle exhaust, and more. This issue has emerged as a substantial global concern, particularly in Thailand, where the levels of PM_2.5 during the summer season have reached hazardous levels. PM_2.5 forecasting is a widely discussed subject that raises awareness and safeguards individuals against pollution. The novelty of this paper is to estimate the weight of linear and nonlinear hybrid models using a differential evolution algorithm. This approach is used for the minimization of the objective function based on hybrid procedures. The data utilized in this study consists of the daily mean PM_2.5 concentration (micrograms per cubic meter) obtained from the Pollution Control Department, Ministry of Natural Resources and Environment, Thailand. The data covers the period from January 2014 to June 2023, encompassing a total of 3,468 observations. Three well-known machine learning approaches, namely the artificial neural network, the long short-term memory, and the convolutional neural network, are employed. We then combined the predicted PM_2.5 obtained from the single machine learning model using linear and nonlinear hybrid procedures. The differential evolution algorithm is utilized to estimate the weight of the hybrid techniques for both scenarios and compare it with state-of-the-art weight approximation. The criteria for evaluating the performance of various hybrid approaches are the performance metrics: the mean absolute error and the median absolute error. The findings of this paper indicate that using a differential evolution algorithm for weight optimization in hybrid procedures outperforms state-of-the-art weight approaches for both linear and nonlinear hybrid models in terms of performance metrics.