An ensemble interval prediction model with change point detection and interval perturbation-based adjustment strategy: A case study of air quality

Abstract

Point prediction has been used to predict air pollutant concentrations in recent years. However, it is still a challenge to characterize the time series data of pollutant concentrations in the presence of high volatility and uncertainty. Since interval prediction can quantity this uncertainty and provide more information than point prediction, we propose an improved interval prediction model based on lower upper bound estimation (LUBE) to construct the prediction intervals (PIs) of PM2.5 concentrations. First, we decompose the original time series into a trend term and a fluctuation term. Then, the trend term with regularity is analyzed by point prediction, while the fluctuation term with uncertainty is studied by LUBE interval prediction. To improve the efficiency and stabilize the randomness in the existing LUBE, we propose a new method based on change point detection and interval perturbation-based adjustment strategy (IPAS). IPAS is used to replace the optimization algorithm in LUBE for improving efficiency. Meanwhile, change point detection is introduced to optimize the initialized parameters in LUBE. In addition, since PM2.5 concentration is influenced by various factors, partial autocorrelation function and maximal information coefficient are applied to select the optimal input features from meteorological and other air pollution factors associated with PM2.5 concentrations. Moreover, we ensemble the prediction results of the trend term and fluctuation term to obtain the ultimate PIs. To evaluate the effectiveness and efficiency of our proposed model, the daily PM2.5 concentration data in Wuhan, China, are analyzed in the empirical study. Comparison results and ablation study clearly indicate that our proposed model has better comprehensive performance than classical models and benchmark models. The proposed new method can provide high-quality PIs and achieve better stability.