Attention Basis Doubly Residual Stacking Neural Network for Multistep-Ahead Prediction in Soft-Sensing Applications

The multistep-ahead prediction is one of the most challenging problems in the field of soft sensing. Traditional time-series prediction methods for soft-sensing applications cannot fully utilize the historical data, neither of them could model the industrial process properly. In this article, the multivariate dynamic deep doubly residual stacking neural network (MD3NN) is first introduced, which is inspired by the neural basis expansion analysis for interpretable time-series forecasting (N-Beats) method. MD3NN incorporates the “doubly residual stacking” idea from N-Beats into a deep stack of fully connected layers for multivariate times-series prediction in soft sensing. Furthermore, a novel spatiotemporal attention basis doubly residual stacking neural network (ADRNN) is proposed. With respect to time and space dimensions of industrial process, ADRNN is designed based on MD3NN to ensure the process being modeled sufficiently. To evaluate the effectiveness of the proposed ADRNN, it is applied to a benchmark debutanizer column (DC) process and a real-world methanation furnace (MF) process. The results illustrate that ADRNN can achieve better prediction accuracy by comparing with other methods.Index Terms— Deep neural network, doubly residual stacking, multistep-ahead prediction, soft sensing, spatiotemporal attention.