Meteorological pattern analysis assisted daily PM2.5 grades prediction using SVM optimized by PSO algorithm

Daily PM2.5 level has significant influence on human health, which is attracting increasing attention. The prediction of PM2.5 grades has thus become an important factor closely related to social development. In the past decades, many prediction methodologies for PM2.5 have been developed, including regression analysis, neural network model, and support vector machine model. Despite these progresses, it still remains a great challenge to predict the PM2.5 grades more accurately and efficiently. In this work, we applied meteorological pattern analysis to assist the support vector machine (SVM) model for PM2.5 class prediction. Cosine similarity was first used to extract three most relevant ones from six common meteorological parameters (atmospheric pressure, relative humidity, air temperature, wind speed, wind direction, cumulative precipitation) to give the needed meteorological pattern for SVM model. Higher prediction accuracy was then obtained with the selected pattern composed by relative humidity, wind speed and wind direction. Moreover, genetic algorithm (GA) and particle swarm optimization (PSO) algorithm were investigated for optimizing the parameters in the process of SVM classification, with PSO-SVM presenting the highest accuracy and efficiency (forecasting time significantly reduced by 25%). We further introduced the criteria of precision, recall and F1-score to evaluate the prediction results of PSO-SVM in each PM2.5 grade. Meanwhile, comparative studies confirmed that PSO-SVM displayed better performance than Adaboost and ANN models for the applied meteorological pattern analysis assisted PM2.5 grades prediction. These obtained results indicate the validity of meteorological pattern analysis for efficient air quality forecasting.