A differentiated DBN model based on CRBM for time series forecasting

Conditional Restricted Boltzmann Machine (CRBM) which adds connections from the last few time-steps to latent and visible variables has been successfully applied for modeling high-dimensional temporal sequences. Considering that the past time-steps observations are more important due to their certainty, in this paper, a differentiated Deep Belief Nets (DBN) model based on CRBM is proposed by introducing two differentiated parameters to improve the importance of past time-steps to achieve higher forecasting accuracy. Based on the proposed model, the energy function of CRBM is reformulated and the update rules of model parameters are induced correspondingly. The experiments verify the influence of differentiated parameters on the forecasting accuracy and illustrate that the proposed model using optimal differentiated parameters can achieve better forecasting performance.