Optimizing neural network architectures for ground temperature prediction in ground source heat pump systems

Ground source heat pump (GSHP) systems typically use vertical borehole heat exchangers (BHEs) as heat sources or sinks. The performance of GSHP systems is highly dependent on the ground temperature surrounding the BHEs, which varies with several parameters during operation. Accurate prediction of this temperature is crucial for optimizing the design and sizing of the BHEs.This study proposes an approach using artificial neural networks (ANNs) to model temperature variations around the borehole based on six operational parameters: time, ground thermal diffusivity, porosity, groundwater speed, heat extraction rate, and distance from the borehole. Various ANN configurations were explored to derive an explicit mathematical equation from the relationships stored within the trained ANN model. The Levenberg–Marquardt algorithm was identified as the most suitable for training, achieving a maximum accuracy of R2 = 0.9997 with ten neurons and the Tansig transfer function in both layers. Simpler configurations, such as an ANN with two neurons (R2 = 0.9783) and one neuron (R2 = 0.9437), were also evaluated to provide more practical equations with reduced complexity.The results demonstrate that careful selection of ANN structure is essential for balancing accuracy and usability. The derived mathematical equations can assist system designers in predicting ground temperature changes for different operational conditions, thus optimizing GSHP system performance. Future work will validate the model with real-world GSHP data to further enhance its practical applicability.