Ground-level ozone prediction using multilayer perceptron trained with an innovative hybrid approach

Abstract

To provide reliable and accurate forecasting on troposphere ozone level is a challenge and significant job related to public heath. Multilayer perceptron (MLP) model trained with original particle swarm optimization (PSO) algorithm, named as MLP-PSO model in this study, can be used to achieve such task, but PSO algorithm tends to suffer from ‘curse of dimensionality’ and overfitting problem, when MLP architecture size becomes large. Such disadvantage inevitably degrades model's predictive performance. Instead of using a random swarm initializing strategy in the MLP-PSO, the hybrid Monte Carlo (HMC) method is employed to sample the weight matrix from the posterior probability distribution of MLP optimal weight matrix, and these sampled weight matrixes are then used to initialize “eight matrix swarm” in PSO before MLP training starts. The MLP-PSO with such new swarm initialization strategy can then be considered as a two-staged hybrid model. Such initial weight matrixes locate in more promising region in PSO search space and hence have better convergence rate and reliability to optimal weight matrix than those by the MLP-PSO with random initialization for analyzing the experiment results. Aforementioned problems encountered by MLP-PSO with random initialization are avoided by our hybrid model. Within our expectation, the experiment results of 1-day ahead forecasting for daily maximum O 3 level in two selected air monitoring sites show that hybrid model has better predictive performance not only in non-episodes but also in episodes.