Potential evaluation of AI-based surrogate models to find the robust and optimal method for thermal performance assessment of desiccant air-conditioning systems

The thermal performance assessment of desiccant air-conditioning (DAC) systems is critical for improving energy efficiency in climate control applications. This study aims to develop an optimal deep learning model to predict the thermal performance of silica gel-based standalone and Maisotsenko cycle (M-cycle) integrated DAC systems. Fourteen different tuning models were constructed using various optimization strategies, including Bayesian optimization (GBRT and GPR), random search, dummy models, and random forest models with various acquisition functions (EI, PI, RCB, and GB hedge). These models were fine-tuned using a comprehensive set of hyper-parameters, such as the learning rate, dense layers, activation function, initialization mode, optimizer, decay rate, batch size, and epochs. The performance of each model was evaluated based on the R² values, with the optimal model achieving an R² value of 0.998, demonstrating high prediction accuracy. This optimal model was trained on an experimental dataset of DAC systems to predict thermal performance. The findings highlight the effectiveness of using AI-based surrogate models for the robust and optimal assessment of DAC systems, contributing to more efficient design and operation strategies for climate control technologies.