Search

• The most important desiccant wheel parameters are considered in the artificial neural network model. • The general trained ANN model predicts DW behavior for arbitrary input parameters. • The volume of the training database is determined by a comprehensive sensitivity analysis. • Various DW ANN models are trained to find the most accurate. • The obtained ANN model predictions are validated with experimental measurements. The design and operational parameters of a solid desiccant wheel (DW), as the heart of a desiccant system, significantly affect the performance of systems in which it is applied. A general and fast model to predict the DW operation is needed to allow investigation of desiccant systems under various working conditions. This paper presents a comprehensive study of the performance prediction of a DW using an artificial neural network (ANN) technique. Comprehensive design parameters, namely rotational speed, channel length, hydraulic diameter, process/regeneration section area ratio, desiccant layer thickness, and the purge angle, are applied as inputs to the ANN model. Moreover, the inlet process and regeneration air temperatures, humidity ratios, and velocities are considered as inputs to confirm that the model outcomes are applicable when various DW parameters are taken into account. Here, the network outputs are the process, the purge, and the regeneration air temperatures and humidity ratios at the DW outlet. This generality in the DW ANN model that contains most of the attributed parameters as inputs and predicts various outputs has not been provided in previous works. A data bank for training the network is generated through transient equations for a wide variation range of DW associated parameters. Several ANN structures are trained in MATLAB to study their prediction accuracy and eventually to find the optimum. The best architecture, which has four hidden layers and 50 neurons, yields a mean relative percentage and a mean square error of 0.54% and 5.9 × 10−5, respectively. This structure can successfully predict DW behavior. Comparing the ANN model predictions and experimental data indicates that the selected structure permits reasonably accurate predictions of the performance of a DW. [ABSTRACT FROM AUTHOR]