Bioinspired Encoder–Decoder Recurrent Neural Network with Attention for Hydroprocessing Unit Modeling

Abstract

The hydroprocessing technique is used to refine crude oil and produce lighter, valuable products. Developing models of these units is crucial for predicting the process dynamics and facilitating optimization and control. In this research, we develop attention-based encoder–decoder recurrent neural network (A-ED-RNN) models, employing various RNN cells such as bioinspired neural circuit policies (NCPs), gated recurrent unit (GRU), and long short-term memory (LSTM), to predict diesel and jet production rates within an industrial hydroprocessing unit. A key innovation is integrating the NCP into the A-ED-RNN models, harnessing its advanced computational power to attain enhanced performance with a smaller model size compared to that of GRU and LSTM cells. The developed RNN models effectively capture the dynamics of diesel and jet production, surpassing the traditional data-driven models. Notably, the NCP-based A-ED-RNN model demonstrates superior memory efficiency and predictive ability, standing out among all of the developed RNN models, underscoring its potential for modeling complex processes.