Time Series Forecasting Using Online Performance-based Ensemble Deep Random Vector Functional Link Neural Network

Abstract

Time series forecasting remains a challenging task in data science while it is of great relevance to decision-making in various industries such as transportation, finance, electricity resource management, meteorology. Traditional forecasting models based on statistics fail in challenging tasks with high non-linearity and complicated characteristics. Due to its architecture bias, deep learning-based models overfit randomness and noise. This paper proposes a novel online performance-based ensemble deep random vector functional link neural network model for the time series forecasting tasks. The proposed model supports the non-iterative online learning and dynamic ensemble method, which keeps adjusting the parameters and the weights of each output layer based on the dynamic evaluation of the latest prediction performance. Extensive experiments show that our proposed method outperforms the state-of-the-art statistical, machine learning-based, and deep learning-based models.