State estimation for industrial desiccant air dryers using hybrid mechanistic and machine learning models

Abstract

Industrial gas drying systems with twin silica gel packed beds are widely used to remove moisture from gas streams, alternating between drying and regeneration phases to maintain continuous operation. This paper presents an integrated solution for real-time estimation of water content within the packed beds of such a system used for drying process air. First, two mechanistic models were developed and validated using gas exit temperature data, achieving average prediction error rates of 5.6% and 5.2% for the drying and regeneration processes, respectively. These mechanistic model prediction errors were then used to train a machine learning model, reducing previous prediction errors to 1.2% and 3.7%. The models were subsequently combined into a hybrid structure and embedded within a moving horizon estimation framework to monitor the internal states of the mechanistic model in real-time. This approach meets observability requirements and provides a more robust solution than open-loop predictions, as demonstrated in the paper through studies using historical process data for both drying and regeneration operations. Further analysis of the regeneration phase revealed that the quantity of heated air and heating duration exceed requirements for water desorption, indicating potential areas for energy optimization using the proposed state estimation solution.