Incorporating first-principles information into hybrid modeling structures: Comparing hybrid semi-parametric models with Physics-Informed Recurrent Neural Networks

Abstract

With increased data availability in the (bio)chemical processing industries, there is a renewed interest in leveraging data-based methods to improve process operations. While data-based approaches enable the modeling of phenomena that are difficult to model mechanistically, they also have drawbacks, especially when presented with serially correlated process data with low variation typically found in the process industries. The limitations in both mechanistic and data-based modeling can be addressed through hybrid approaches. The combination of mechanistic and data-based models into hybrid semi-parametric models has shown great promise in mitigating such limitations over the last 30 years. More recently, physics-informed learning approaches have been proposed as an alternative method for embedding process knowledge in data-based models. This work provides a comparative study of hybrid semi-parametric modeling and Physics-Informed Recurrent Neural Networks (PIRNNs) applied to a pilot-scale bubble column aeration case study. The developed models are compared based on the ease of training, the models’ adherence to the governing system equations, the prediction accuracy when reducing the measurement frequency, and the model performance when reducing the quantity of training data. For the considered case study, the hybrid semi-parametric modeling approach generally resulted in superior model performance with high prediction accuracy, good adherence to the physics, and good performance when reducing the quantity of training data.