(372z) Time Series Modelling of an Industrial Desiccant Dryer

In several industrial processes and applications where the presence of moisture can lead to problems such as corrosion, product degradation, inaccurate measurements, or process inefficiency, bone dry air, which is air with an extremely low moisture content, is crucial. For large-scale processes with over 10,000 m³/hr air flows, packed bed desiccant dryers are the preferred unit operation used to produce bone dry air. Typically, these dryers consist of two pressure vessels filled with silica gel used to adsorb water vapor from the air. When one vessel reaches its maximum adsorption capacity, the other vessel takes over the operation while the saturated vessel undergoes regeneration. This cyclic process necessitates predictive modelling to manage it effectively. Therefore, accurately modelling the discharge temperature of the dryer is essential, especially since monitoring the temperature within the reactor is not feasible. Moreover, unstable cycles may lead to operational shutdowns, resulting in revenue loss. However, the complexity of modelling adsorption phenomena presents challenges in developing mechanistic models for desiccant dryers.

Addressing these challenges, this study proposes a novel approach by integrating mechanistic models with machine learning techniques, utilizing a hybrid superposition approach. Acknowledging the difficulty in directly measuring internal states of the dryer, such as the extent of adsorption inside the bed — critical for understanding and predicting the drying process— we leverage moving horizon estimation (MHE). MHE can infer the internal states from measurements of temperature and dew point of air exiting the process, thus offering a more complete picture of the system's dynamics.

Mass and energy balance - empirical adsorption rate constant - compartmentalized modelling multiple layers in series to capture spatial dependence. Model parameters estimated by solving nonlinear least squares problem using a surrogate optimization approach - where the number of compartments is rounded to the nearest integer but treated as a continuous parameter.

The data-driven component of our study involves an LSTM structure tasked with forecasting the discharge temperature of the dryer, a crucial indicator of operational efficiency. Input features include total and dilution air mass flowrates, dryer inlet and outlet pressures, and air inlet temperature. Furthermore, we utilize Optuna (Akiba et al., 2019) for the optimization of the LSTM network's architecture, determining the optimal number of layers and neurons to enhance prediction accuracy. Our findings, within the explored range of 2-8 hidden layers and 10-100 neurons, identify an optimal configuration of 5 hidden layers and 96 neurons.

By combining the predictive power of LSTM with the insights derived from mechanistic models and MHE, this study sets a precedent for a more informed and adaptive approach to managing packed bed desiccant dryers. This integrated model not only promises to improve the reliability and efficiency of these critical systems but also mitigates the risk of operational disruptions, paving the way for more sustainable and cost-effective industrial processes.