An Improved Parallel MDBN With AVMD for Nonlinear System Modeling

Abstract

Nonlinear system modeling using Deep Belief Network (DBN) is currently a research hotspot. However, the training process of DBN needs large amount of data to guarantee accuracy, and the traditional DBN may not meet the requirement of high-precision modeling. In this paper, we first improve the original DBN and Variational Mode Decomposition (VMD) algorithms, and on this basis, we then proposed a parallel Momentum Deep Belief Networks (MDBN) with Adaptive Variational Mode Decomposition (AVMD). Parallel AVMD-MDBN is an improved modeling method based on the deep learning model DBN. Firstly, a single raw dataset is decomposed into a specific number of sub-datasets using AVMD. Then these sub-datasets are distributed among a number of improved MDBNs. A single raw dataset learning model and algorithm is extended to multiple feature extraction nodes to learn the characteristics of multiple sub-datasets in parallel. Finally, the results of the multiple nodes are transmitted to the main feature extraction node to complete the regression calculation. In order to verify the effectiveness of the model, the proposed parallel AVMD-MDBN model is tested on a nonlinear dynamic system modeling, a Mackey-Glass time-series prediction and a financial stock prediction. Our experimental results show that the proposed parallel AVMD-MDBN has better performances in terms of improving feature learning ability than that of other methods.

Keywords

• Parallel momentum deep belief networks
• adaptive variational mode decomposition
• contrast divergence algorithm
• nonlinear system modeling
• financial stock prediction