Data-Driven Dynamic Process Modeling Using Temporal RNN Incorporating Output Variable Autocorrelation and Stacked Autoencoder

Dynamic process modeling in process industries has been extensively studied, especially with the development of deep learning techniques. Recurrent neural networks (RNN) and stacked autoencoders (SAE) are two powerful tools for dynamic modeling and data processing. However, most existing research primarily focuses on extracting features from process input data, often neglecting the temporal autocorrelation of output variables. In this work, a hierarchical model based on time-series RNN structure is proposed. The upper layer employs a long short-term memory (LSTM) network to extract temporal features from process input data. The lower layer uses a gated recurrent unit (GRU) to model the temporal dependencies of output variables across samples. These two parts are concatenated to form the model. Additionally, SAE is utilized to perform dimensionality reduction and reconstruction of process input, seamlessly integrating the reconstruction process with the RNN into a unified framework, termed the AR-SAE-RNN model. Distributed training is employed to effectively learn the model parameters. The proposed AR-SAE-RNN model has been applied to an industrial distillation column. The results demonstrate the model's effectiveness in capturing the dynamic behavior of the system.