Yield and Properties Prediction Based on the Multicondition LSTM Model for the Solvent Deasphalting Process.

Area of Science:

• Chemical Engineering
• Process Control
• Data-Driven Modeling

Background:

• Solvent deasphalting (SDA) is a critical industrial process with complex, multiscale dynamics.
• Existing data-driven models often overlook the impact of varying operation modes on process outcomes.
• Accurate prediction and control are essential for optimizing yield and ensuring operational safety.

Purpose of the Study:

• To develop a data-driven model that accounts for multiple operation modes in SDA processes.
• To improve the prediction accuracy of key process variables like DAO yield and Conradson carbon residual.
• To enhance fault detection capabilities by incorporating operational context.

Main Methods:

• Feature selection using random forests to identify critical input variables.
• Stack denoising autoencoder (SDAE) for data reconstruction, feature dimension reduction, and improved clustering.
• Fuzzy c-means clustering to accurately divide SDA process into distinct operation modes.
• Development of a multicondition Long Short-Term Memory (LSTM) model for time-lag process prediction.

Main Results:

• Feature selection identified key variables influencing DAO yield and Conradson carbon residual.
• SDAE effectively reduced feature dimensions and improved operation mode classification accuracy.
• The multicondition LSTM model achieved prediction accuracy with R² > 0.95, outperforming traditional LSTM.
• Sensitivity analyses confirmed consistency with two-phase countercurrent extraction principles.
• The method demonstrated effectiveness in fault detection on the Tennessee Eastman process benchmark.

Conclusions:

• The proposed multicondition LSTM model accurately predicts SDA process behavior across different operation modes.
• This approach significantly enhances prediction accuracy and enables earlier, more reliable fault detection.
• The methodology is transferable to other complex industrial processes, offering broader applicability.