1 3 Heat Mass Transfer DOI 10.1007/s00231-015-1723-z ORIGINAL Performance prediction between horizontal and vertical source heat pump systems for greenhouse heating with the use of artificial neural networks Hüseyin Benli 1 Received: 6 January 2015 / Accepted: 17 November 2015 © Springer-Verlag Berlin Heidelberg 2015 CGP Pola–Ribiere conjugate gradient learning algorithm cov Coefficient of variation (%) LM Levenberg–Marquardt learning algorithm n Number of independent data patterns RMS Root-mean square error R 2 Fraction of variance SCG Scaled conjugate gradient learning algorithm t Target T o Outdoor air temperature (°C) T i Indoor air temperature (°C) T g1 Temperature of ground at 2 m depth (°C) T g2 Temperature of ground at 60 m depth (°C) T o,wa Outlet average water-antifreeze solution tempera- ture of HGHE or VGHE (°C) T i,wa Inlet average water-antifreeze solution tempera- ture of HGHE or VGHE (°C) T o,air Average air temperature leaving condenser fan- coil unit (°C) T i,air Average air temperature entering condenser fan- coil unit (°C) y Calculated neural network output Subscripts g 1 Ground at 2 m depth g 2 Ground at 60 m depth H Horizontal HP Heat pump i Inlet (inside) o Outlet (outside) sys System V Vertical mea Measured pre Predicted Abstract This paper presents the suitability of artificial neural networks (ANNs) to predict the performance and comparison between a horizontal and a vertical ground source heat pump system. Performance forecasting is the precondition for the optimal control and energy saving operation of heat pump systems. In this study, performance parameters such as air temperature entering condenser fan- coil unit, air temperature leaving condenser fan-coil unit, and ground temperatures (2 and 60 m) obtained experi- mental studies are input data; coefficient of performance of system (COP sys ) is in output layer. The back propaga- tion learning algorithm with three different variants such as Levenberg–Marguardt, Pola–Ribiere conjugate gradient, and scaled conjugate gradient, and also tangent sigmoid transfer function were used in the network so that the best approach can be found. The results showed that LM with three neurons in the hidden layer is the most suitable algo- rithm with maximum correlation coefficients R 2 of 0.999, minimum root mean square RMS value and low coef- ficient variance COV. The reported results confirmed that the use of ANN for performance prediction of COP sys,H–V is acceptable in these studies. List of symbols ANN Artificial neural network COP sys Heating coefficient of performance of ground- source heat pump system * Hüseyin Benli hbenli@firat.edu.tr; hbenli@msn.com 1 Department of Technical and Vocational Education, Fırat University, 23119 Elazig, Turkey