comparing between predicted output temperature of flat-plate solar collector and experimental results: computational fluid dynamics and artificial neural network

introduction the significant of solar energy as a renewable energy source, clean and without damage to the environment, for the production of electricity and heat is of great importance. furthermore, due to the oil crisis as well as reducing the cost of home heating by 70%, solar energy in the past two decades has been a favorite of many researchers. solar collectors are devices for collecting solar radiant energy through which this energy is converted into heat and then heat is transferred to a fluid (usually air or water). therefore, a key component in performance improvement of solar heating system is a solar collector optimization under different testing conditions. however, estimation of output parameters under different testing conditions is costly, time consuming and mostly impossible. as a result, smart use of neural networks as well as cfd (computational fluid dynamics) to predict the properties with which desired output would have been acquired is valuable. to the best of our knowledge, there are no any studies that compare experimental results with cfd and ann. materials and methods a corrugated galvanized iron sheet of 2 m length, 1 m wide and 0.5 mm in thickness was used as an absorber plate for absorbing the incident solar radiation (fig. 1 and 2). corrugations in absorber were caused turbulent air and improved heat transfer coefficient. computational fluid dynamics k-ε turbulence model was used for simulation. the following assumptions are made in the analysis. (1) air is a continuous medium and incompressible. (2) the flow is steady and possesses have turbulent flow characteristics, due to the high velocity of flow. (3) the thermal-physical properties of the absorber sheet and the absorber tube are constant with respect to the operating temperature. (4) the bottom side of the absorber tube and the absorber plate are assumed to be adiabatic. artificial neural network in this research a one-hidden-layer feed-forward network based on the back propagation learning rule was used to simulate the output temperature of a solar collector. the number of neurons within the hidden layer varied from 1 to 20. the hyperbolic tan- sigmoid and pure-line were used as the transfer function in the hidden layer and output layer, respectively. minimization of error was achieved using the levenberg-marquardt algorithm. to carry out the aforementioned steps, the dataset (105 observations) was split into training (70 observations), and test (35 observations) data. training sets used to develop models included air velocity, solar radiation, time of the day, ambient moisture and temperature values as inputs with an associated temperature of the collector as outputs. the aim of every training algorithm is to reduce this global error by adjusting the weights and biases. results and discussion compare experimental results with ann the performance of the three-layer ann for the prediction of output temperature of flat-plate solar collector by the levenberg–marquardt training algorithm was illustrated in fig. 4. ann predicted output temperatures with r2 and rmse of 0.92 and 1.23, respectively. furthermore, the maximum error in prediction of output temperature of solar collector was 3.3 k. these results are in agreement with tripathy and kumar, (2009) those who have predicted the output temperatures of food product in the solar drier using ann with and rmse of 0.95 and 0.77, respectively. compare experimental results with cfd simulation fig. 6 shows that over the starting length of the absorber plate, there is a variation of the velocity profile which is caused by sharp geometry and it leads to some recirculation of the air in this part of absorber plate. after this part of boundary layers, flow is fully developed and velocity profile becomes smoother and constant. fig. 8 shows that the predicted temperature was within the experimentally measured temperature. the highest differences between simulated and experimental temperatures were around -2.4k to 4.6k for different time periods. the temperature differences of 4k were reported by selmi et al. (2008). this disagreement is due possibly to the fact that there are unknown experimental inputs such as turbulence intensity, radiative heat loss from the absorber sheet to the surroundings, leakage, and measurement tool errors which were not accounted in the model simulations. these losses by radiation are significant at high irradiation levels. this result agrees with studies done in badache et al. (2012). thickness of absorber plate and radiation loss, in cfd model, does not take into consideration. for this reason maximum output temperature is seen in maximum radiation which is 12 p.m. while in real condition, it takes some time for absorber plate to get to its maximum temperature. moreover, the numerical temperature is smaller than the real temperature after 12 p.m. this may occur because of the thickness of metal which keeping the absorbed heat and losing it after awhile. generally there is a time step hysteresis for the numerical temperature. conclusions according to this study it can be concluded that the ann operates better than cfd to predict the output temperature operation. however, ann method does not give any information about the prediction of temperature distribution and velocity profiles in the solar collector. although prediction accuracy of the cfd method is less than ann method, but the provided information on the velocity and temperature profile of the solar collector is still valuable.