Prediction models of air outlet states of desiccant wheels using multiple regression and artificial neural network methods based on criterion numbers

Abstract

• Study simple and general models to calculate air outlet states of desiccant wheels. • Get 3 basic dimensionless numbers through nondimensionalizing physical equations. • Calculate 2500 groups original data with the nondimensionalized mathematic model. • Errors of the multiple regression formulars are from ±5% to ±15%. • Errors of the model using BP artificial neural network are from ±5% to ±10%. In this study, multiple regression and artificial neural network methods were applied to desiccant wheels to calculate the dimensionless outlet temperatures and humidity ratios of the processed air (A pro ) and regeneration air (A reg ). First, coupled heat and mass transfer equations expressing the relationship between the air and desiccant wheel were nondimensionalized to obtain seven criterion numbers, including three basic criterion numbers. The three basic criterion numbers were related to the structural and operating parameters of desiccant wheels and the physical parameters of the air and desiccants. Subsequently, the value ranges of the three basic criterion numbers were obtained based on the ranges of the above parameters. Based on a wide range of air inlet conditions and value ranges of the three basic criterion numbers, 2500 cases were designed. The nondimensionalized equations were used to calculate the dimensionless outlet temperatures and humidity ratios of A pro and A reg for the 2500 cases. Next, a multiple regression method was used to regress formulas for a fixed air inlet condition based on the original calculated results. The formulas were related to the three basic criterion numbers and could be used to calculate the dimensionless outlet temperatures and humidity ratios of A pro and A reg . Compared with the original data, the prediction errors of the formulas ranged from ±5% to ±15% for the analyzed working conditions. Finally, based on the 2500 original data groups, the backpropagation neural network was used to predict the dimensionless outlet states of A pro and A reg , with the three basic criterion numbers and air inlet conditions as the inputs. Compared with the original data, the prediction errors of the four dimensionless outputs mainly ranged from ±5% to ±10%. The difference between the two prediction methods was within ±10%.