Time series prediction using dynamic Bayesian network

Time series prediction is a challenging research topic, especially for multi-step-ahead prediction. In this paper, a novel multi-step-ahead time series prediction model is proposed based on combination of the Kalman filtering model (KFM) and the echo neural networks (ESN). Recently, the studies demonstrate the ESN model is a promising strategy for multistep-ahead time series prediction, at the same time, the KFM is a recursion-based sequence information processing approach, which has been used effectively for prediction, filtering and smooth of time series data. In this paper, we consider to use the recursion-based KFM to enhance performance of the ESN-based direct prediction model. A novel graph model named the E-KFM that generated from combination of the ESN and the KFM is developed to predict multi-step-ahead time series data. The simulation and comparison results show that the proposed model is more effectiveness and robustness. (C) 2017 Elsevier GmbH. All rights reserved.