Novel Spatiotemporal NBEATS for Long-term Industrial Soft Sensor

Forecasting long-term industrial process data is challenging due to its inherent characteristics of high nonlinearity, dynamic and noises. Many existing industrial soft-sensor modeling methods enable the prediction of dominant variables based on easily measurable process variables but cannot predict the values of dominant variables over long time horizons. Therefore, a novel spatiotemporal NBEATS (ST-NBEATS) model is proposed for forecasting long-term dominant variables in industrial processes. Trend stack and covariate stack are designed to capture intrinsic long-term characteristics. Then the graph convolutional network (GCN) is incorporated to determine the dynamic mutual interactions. To illustrate the effectiveness and superiority of the proposed method, the ST-NBEATS is evaluated on the Tennessee Eastman Process (TEP) case study. The experiment results show that ST-NBEATS achieves state-of-the-art compared to the recurrent neural network (RNN), the long short term memory (LSTM), the gated recurrent unit (GRU) and the probabilistic slow feature analysis (PSFA).