Typically, g-functions are used to simplify and speed up the simulation of ground heat exchangers, but they usually neglect the layered heterogeneity and groundwater flow present in most natural environments. In this work, an artificial neural network capable of approximating the short-term transfer function at the outlet fluid temperature of a closed-loop borehole in a stratified hydrogeological environment is proposed. A total of 40,000 simulations were completed with a finite element model to construct the database used for training and testing. The accuracy of the neural network is measured on an independent test subset of 3,400 simulations with an average root mean square error of 5.4 Ã— 10â»Â³ Â°C. A new architecture, based on the combination of several neural networks, is also put forward allowing greater latitude in the approximation of complex nonlinear phenomena and likely reducing the drawbacks associated with the training database size.

Short-term transfer function approximation using artificial neural network in a heterogeneous medium with groundwater advection