An integrated framework for plant data-driven process modeling using deep-learning with Monte-Carlo simulations

This study aims to develop a deep-learning-based and plant data-driven framework for process modeling to help understanding plant-wide processes. The systematic framework consists of the following steps: data processing based on domain-knowledge, deep-learning model development, model selection using information criteria, and global sensitivity analysis with Monte-Carlo simulations. The assessment of the quality of the optimal deep-learning model to support plant-wide process understanding is the key emphasis of this framework. The proposed framework was applied for analyzing long-term data from wastewater treatment plants to predict nitrous oxide emission characteristics. The results showed a promising potential of the framework to systematically and efficiently develop fit-for-purpose deep-learning models with highly favorable cross-validation statistics (R²). The framework is expected to facilitate the development of versatile deep-learning models based on plant data encompassing nonlinear and complex process phenomena, where especially mechanistic models are not available.