Efficient g-function approximation with artificial neural networks for a varying number of boreholes on a regular or irregular layout

Abstract

A g-function links heat extraction rates to the mean borehole wall temperatures in geothermal heat exchangers. New computation methods are proposed continuously, becoming more accurate as they rise in complexity. In this work, it is shown how artificial neural networks can be assembled and trained to approximate g-functions rapidly and accurately for varying numbers of boreholes placed on a regular or irregular layout. This article focuses on geothermal heat exchangers made up of one to ten boreholes in a field of 30 by 30 m with various borehole geometries and soil thermal properties. After a training phase that used a database of 500,000 g-functions obtained with a block matrix formulation, a mean computation time of 11.18 ms was reached. This is 37% faster than using the block matrix formulation and represents a small fraction of the time required to simulate the fluid temperature over a 5-year period. The accuracy of the network against regular and irregular configurations of boreholes was evaluated and median errors of 0.18% and 0.39% were noted. The methodology introduced in this work is general and can be used together with any method allowing construction of g-functions.