Application of artificial neural networks to near-instant construction of short-term g-functions

• Published in: Applied thermal engineering Vol. 143; pp. 910 - 921
• Main Authors: Pasquier, Philippe, Zarrella, Angelo, Labib, Richard
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.10.2018; Elsevier BV
• Subjects: Architecture; Artificial neural network; Artificial neural networks; Boreholes; Boundary conditions; Computer simulation; Cooling; Ground heat exchanger simulation; Ground-coupled heat pump system; Heat exchangers; Heating; Iterative methods; Mathematical models; Model accuracy; Multilayer perceptrons; Neural networks; Short-term g-function; Source code; Thermal resistance; Thermal resistance and capacity model; Training; Artificial neural network; Ground-coupled heat pump system; Short-term g-function; Ground heat exchanger simulation; Thermal resistance and capacity model
• ISSN: 1359-4311; 1873-5606
• DOI: 10.1016/j.applthermaleng.2018.07.137

•An artificial neural network accurately constructs a short-term g-function.•A training set of 15 000 g-functions is assembled with a lumped element model.•The trained network is validated with numerical and experimental data sets.•The short-term g-function is accurately constructed in a few milliseconds.•A source code combining a short- and long-term g-function is provided. The concept of a g-function is frequently used to simplify the simulation of a ground heat exchanger composed of several boreholes. The technique is suitable for modelling long-term thermal interactions, although it suffers from a lack of accuracy in the short term. Alternatively, models based on the theory of thermal resistances and capacities provide a more accurate solution for the short term but at the cost of a long solution time that prevents an efficient iterative use. In this work, an artificial neural network designed to rapidly and accurately construct the short-term g-function of a ground heat exchanger is presented. To assemble a large training set, 15 000 simulations are performed with an accurate thermal resistance and capacity model. The boundary conditions used to solve the model allow direct construction of the short-term g-function for the outlet fluid temperature. A multilayer perceptron architecture is proposed in conjunction with a methodology designed to reduce training time and increase the accuracy of the neural network. The output of the resulting network is compared with numerical and experimental data sets. The construction of a short-term g-function by an artificial neural network is 10 000 times faster and achieved in a few milliseconds. A source code combining efficiently a short- and long-term g-function is published with the paper.