A stacked ensemble learning method for traffic speed forecasting using empirical mode decomposition

Abstract

ABSTRACT Traffic speed is an important traffic parameter whose accurate prediction is crucial in traffic management. Different machine learning and ensemble methods have been used for predicting traffic parameters. Proper feature selection, the type of learners, ensemble methods, and appropriate aggregation technique have a significant impact on the prediction accuracy of ensemble methods. ‎To achieve higher prediction accuracy, using the nonlinear learning ability of Support Vector Regression (SVR) model, we propose a stacking ensemble learning model.‎‏‎ We train the base learners using different training subsets based on random subspace selection. In the analysis of empirical and complex data, Empirical Mode Decomposition (EMD) is used for decomposing time series data into seasonal and trend components. We use EMD to create new features. In order to reduce the computation time, we also propose a sample selection method. The results of our experiments show that the proposed method outperforms the existing hybrid EMD-ARIMA and non-ensemble models. In comparison with the baseline models, using the EMD technique leads to a 0.32 to 4% decrease in RMSE for different links. The results show that the proposed model outperforms the ARIMA and ARIMA-EMD models in terms of long-term traffic speed forecasting performance.