(346j) A Hybrid Mechanistic-Machine Learning Approach to Identify Dynamical Models for Sustainability Assessment of Manufacturing Processes: A Soybean Diesel Process Case Study

Current dynamic models for full industrial process plants exist as highly accurate mechanistic models based on first-principle relationships that can be quickly found using the dynamic simulation capabilities of a chemical process model software like Aspen Plus Dynamics. However, the prohibitively high order of these models limits analysis of physical interpretation of terms or applications where computational simplicity is a requirement. For these types of applications, lower-order dynamic models that may sacrifice accuracy for simplicity prove more useful. This type of low-order dynamic model is especially promising as a method for analysis of system sustainability by revealing dynamic relationships between critical mass and energy flows. Despite this, there have been few attempts at finding these low-order models of chemical manufacturing processes, with existing work focusing on model development of individual components and weak connections to long term sustainability. This work seeks to fill this gap in the literature by developing a reduced-order model using a hybrid mechanistic-machine learning approach and demonstrating the approach on an entire soybean-oil to soybean-diesel process plant. We use a grey-box machine learning method with a standard nonlinear optimization approach to identify relevant models of governing dynamics as ODEs using the data simulated from mechanistic models in ASPEN Plus Dynamics. Results show that the method identifies a linear ordinary differential equation model that gives an accurate relation to output, input, and selected internal molar flow rates. The identification of a linear model is reflective of underlying stoichiometric mechanisms driving the dynamics for this particular system. We further analyze the water-usage sustainability of this process by identifying model terms relating the system's soybean-diesel output to water consumption, and consider the implications of this interaction if the demand for soybean-diesel increases. This work demonstrates the strength of reduced order model development for long term sustainability assessment of manufacturing processes with lower computational efforts.