Predicting Hourly $\mathbf{PM}_{2.5}$ Concentrations Based on Random Forest and Ensemble Neural Network

This study presents a method to predict hourly PM2.5 concentrations by combining random forests and ensemble neural network. Random forest is used to rank variable importance and complete variable selection using out-of-bag error. Aiming to increase prediction accuracy and robustness of the neural network model, ensemble neural network using selected input variables is developed as the prediction model, whose diversity is increased by randomly choosing the hidden neuron number of each learner from a certain set. Moreover, the pattern between PM2.5 concentrations and input variables are changing with time. In order to capture well new patterns, new samples are given larger sampling probabilities in the bagging. Experimental results demonstrate that the ensemble neural network has better performance than the persistence and the random forest in terms of accuracy, and giving the larger sampling probabilities to last samples improves the prediction accuracy of ensemble neural network.