Abstract
Data-driven soft sensors have been widely applied in complex industrial processes. However, the interpretable spatio-temporal features extraction by soft sensors remains a challenge. In this light, this work introduces a novel method termed spatio-temporal consistent and interpretable model (STCIM). First, temporal and spatial features are captured and aligned by a far topological spatio-temporal consistency extraction block. Then, the features are mapped into an interpretable latent space for further prediction by explicitly giving physical meanings to latent variables. The efficacy of the proposed STCIM is demonstrated through the modeling of two generated datasets and a real-life dataset of coal-fired power plants. The corresponding experiments show: 1) The generalization of STCIM outperforms other methods, especially in different operation situations. 2) The far topological spatio-temporal consistency is vital for feature alignment. 3) The hyper-parameters of physics-informed interpretable latent space loss decide the performance of STCIM.