Air Pollutant Concentration Forecast Based on Support Vector Regression and Quantum-Behaved Particle Swarm Optimization

Abstract

In order to improve the forecasting accuracy of atmospheric pollutant concentration, a prediction model of atmospheric PM2.5 and nitrogen dioxide (NO2) concentration based on support vector regression (SVR) is established. Quantum-behaved particle swarm optimization (QPSO) algorithm is used to select the optimal parameters influencing the performance of SVR. And in order to improve the problem that the fixed SVR model is difficult to adapt to the highly nonlinear process, a simple online SVR based on re-modeling method is proposed instead of the fixed one. According to hourly PM2.5 and NO2 concentrations and meteorological conditions from May 2014 to April 2015 in Wanliu Monitoring Station of Beijing in China, the experiment is carried out based on the data of 3 months. Meanwhile, PM2.5 concentration is predicted by three different prediction methods, including the recursive prediction method, direct prediction method, and online direct prediction method. The results show that the online direct prediction method is the most accurate in the three prediction methods. In addition, compared with original particle swarm optimization (PSO) algorithm, QPSO algorithm is tested more efficiently for the improvement of global search ability and robustness during the procedure of parameter selection. Moreover, the hybrid QPSO-SVR model proposed in this paper has higher prediction accuracy and less computational time compared with the PSO-SVR model, genetic algorithm (GA)-SVR model, and grid search (GS)-SVR model, which indicates reliability and effectiveness of the QPSO-SVR model in prediction of these two pollutant concentrations.