The porous volumetric receivers are used widely in concentrated solar systems due to their ability to withstand high heat fluxes, elevated working temperatures and corrosive environments. Since the performance of these receivers depends on a number of structural and design factors, optimization studies are necessary to get the greatest performance and create better receiver designs. Optimization studies which use direct numerical simulations are generally time-consuming and computationally expensive. The current study investigates the potential of Artificial Neural Networks (ANN) for faster and more accurate thermo-hydraulic performance prediction of a porous volumetric solar receiver, which may also be utilized to replace numerical simulations during optimization studies. Several variations of training algorithms have been developed for updating weights and biases of the ANN model. These training algorithms vary in accuracy, speed and computational requirements depending on the type of problem and complexity of the input-output relationship of the training data. Hence, a comparison of various training algorithms in terms of prediction error and training performance is presented in the current work. The prediction accuracy of various training models is compared based on mean absolute percentage error. The training performance is compared based on epochs or iterations and the time required for model training. The porosity, pore size, absorber length and inlet velocity are selected as the input variables for the ANN model in the present work. The output variables for the ANN model are the pressure drop and outlet fluid temperature, indicating the receiver's hydraulic and thermal performances. The training data set is created with various random sets of input and corresponding output values. The corresponding pressure drop and air outlet temperatures are obtained by performing coupled numerical simulations for the porous receiver. The findings show that the Levenberg-Marquardt method provides the best compromise between prediction accuracy and training time and should be used as the default approach for receiver performance prediction. Additionally, it is determined that ANN architecture with three or more layers and at least 20 neurons is necessary to achieve adequate prediction accuracy.