A construction approach to prediction intervals based on bootstrap and deep belief network

Abstract

Traditional neural networks (NNs) have been widely used in prediction intervals (PIs) construction method, with many improved models have been proposed. However, there are not satisfactory prediction results when dealing with some complex prediction problems which have many relative influence factors of the target value and high noise because of the complexity. To get prediction value effectively, a novel method proposed for the construction of PIs with deep belief networks (DBNs) and bootstrap. We apply the DBNs with many hidden layers to predict data. In addition, a cost function based on PIs is introduced into the prediction model which based on DBN to get a better network. Bootstrap technique is used to simplify the process of construction of PIs and improve the ability of anti-interference. Finally, we make experiments in five synthetic and real cases to point the performance of the proposed method. The results show the effectiveness of our method in improving the quality of PIs, especially for some complex problems.