Development of Steady-State and Dynamic Mass and Energy Constrained Neural Networks for Distributed Chemical Systems...

Area of Science:

• Chemical Engineering
• Artificial Intelligence
• Process Systems Engineering

Background:

• Accurate modeling of chemical processes is crucial for efficiency and safety.
• Existing neural network models often struggle to strictly enforce physical laws like mass and energy conservation, especially with noisy data.
• Data-driven approaches offer flexibility but require robust methods to guarantee physical constraints.

Purpose of the Study:

• To develop novel algorithms for mass and energy constrained neural network models.
• To ensure exact conservation of mass and energy in distributed chemical process systems.
• To provide a guaranteed method for satisfying conservation laws, unlike soft penalization techniques.

Main Methods:

• Development of algorithms for equality-constrained nonlinear optimization problems.
• Leveraging hybrid series and parallel dynamic-static neural networks for distributed systems.
• Validation using steady-state and dynamic data with diverse noise characteristics.

Main Results:

• Algorithms successfully guarantee exact mass and energy conservation for dynamic chemical processes.
• Achieved root mean squared error below 1% across various case studies.
• Demonstrated flexibility with system holdup information for dynamic processes.

Conclusions:

• The developed data-driven algorithms provide a robust framework for accurate and physically constrained modeling of chemical processes.
• Mass-energy constrained neural networks offer a significant improvement over traditional methods for dynamic system modeling.
• The approach is applicable to various chemical engineering systems, including reactors and heat exchangers.