Application of machine learning algorithms for power output prediction in thermo-active diaphragm walls
File(s)
Published version
Author(s)
Sanchez Fernandez, Javier
Taborda, David MG
Ruiz Lopez, Agustin
Type
Conference Paper
Abstract
Thermo-active diaphragm walls offer an efficient route to low carbon heating and cooling by embedding heat-exchanging loops within deep retaining structures. In this study, a surrogate model to predict the power output per unit area of a diaphragm wall section is presented and evaluated. A parametrised 3D finite element (FE) model was created in COMSOL Multiphysics and used to assess 500 combinations of wall width and depth, concrete and soil conductivities, inlet–ground temperature differential, convective heat transfer coefficient at the exposed boundary, and fluid velocity. This was used as a numerical database for the training and testing of the surrogate model. The dataset covers 16 time instants retrieved from the transient analysis, in addition to the steady-state solution. The artificial neural network (ANN) regressor subsequently trained achieved an average coefficient of determination (R²) of 0.987. The average error in power prediction is limited to a few W/m², with the model being able to capture the correct trends even in the less accurate samples. Feature importance was assessed using SHAP analysis. This revealed that, as expected, early-time performance is dominated by the temperature differential and concrete conductivity, while the convective heat transfer coefficient at the exposed boundary governs long-term output. By replacing computationally expensive FE runs with fast ANN predictions, the proposed surrogate enables rapid design optimisation, sensitivity studies, and feasibility assessments for a large set of configurations.
Date Issued
2026-06-14
Date Acceptance
2026-05-01
Citation
Proceedings of the 21st ICSMGE, 2026, pp.6033-6036
ISBN
978-3-9503898-4-5
Publisher
ÖGG, Austrian Society for Geomechanics
Start Page
6033
End Page
6036
Journal / Book Title
Proceedings of the 21st ICSMGE
Copyright Statement
© 2026 The Author(s).
Source
21st International Conference on Soil Mechanics and Geotechnical Engineering (ICSMGE)
Publication Status
Published
Start Date
2026-06-14
Finish Date
2026-06-19
Coverage Spatial
Vienna, Austria