Long-term performance prediction of forced circulation solar domestic water heating systems using artificial neural networks

Abstract

The objective of this work is to use Artificial Neural Networks (ANNs) for the long-term performance prediction of forced circulation type solar domestic water heating (SDWH) systems. ANNs have been used in diverse applications and they have been shown to be particularly useful in system modelling and for system identification. Three SDWH systems have been tested and modelled according to the procedures outlined in the standard ISO 9459-2 at three locations in Greece. Two ANNs have been trained using the monthly data produced by the modelling program supplied with the standard. Different networks were used due to the different natures of the output required in each case. The first network was trained to estimate the solar energy output of the system for a draw-off quantity equal to the storage tank capacity and the second network was trained to estimate the solar energy output of the system and the average quantity of hot water per month, at demand temperatures of 35 and 40°C. The data presented as input to both networks are similar to the data used in the program supplied with the standard. The statistical coefficient of multiple determination (R2-value) obtained for the training data set was equal to 0.9972 for the first network and equal to 0.9878 and 0.9973 for the second network for the two output parameters, solar energy output and hot water quantity, respectively. Other data, unknown to the network, were subsequently used to evaluate the accuracy of the prediction. Predictions with R2-values equal to 0.9945 for the first network and 0.9825 and 0.9910 for the second were obtained. The maximum percentage differences were 1.9 and 5.5% for the two networks respectively. These results indicate that the proposed method can successfully be used for the prediction of the long-term performance of forced circulation water heating solar systems. The advantages of this approach compared to the conventional algorithmic methods are speed, simplicity, and the capacity of the network to learn from examples. This is done by embedding experiential knowledge in the network.