A novel machine learning framework based on proper orthogonal decomposition for turbine wake flow predictions in natural waterways
Abstract
We developed a machine learning (ML) surrogate trained on instantaneous large eddy simulation (LES) data filtered using proper orthogonal decomposition (POD) to predict mean flow fields in natural rivers with complex geometries. The surrogate enables efficient prediction of mean flow fields downstream of marine hydrokinetic (MHK) turbine arrays at a fraction of the computational cost of high fidelity simulations, facilitating turbine layout optimization in natural waterways where bathymetry, turbulence, and turbine interactions influence the available energy resource. We demonstrate the capabilities of the model at a real world tidal river site under ebb tide conditions. Complex bed bathymetry and planform geometry generate substantial background turbulence that can obscure turbine wake signatures. POD is employed to extract relevant flow structures, improving the ability of the ML model to learn wake dynamics. First, LES is performed at the tidal site to generate training data. POD is then applied to decompose the instantaneous LES fields into spatial modes. Selected modes are used to train a three dimensional convolutional encoder decoder network to predict time averaged velocity fields. This selective filtering reduces the influence of dominant background fluctuations while preserving information relevant to turbine wakes. At equivalent network size and training cost, the POD filtered surrogate reduces the mean squared error by three orders of magnitude compared with the same network trained on unfiltered LES data.
Related
- Extended proper orthogonal decomposition of non-homogeneous thermal fields in a turbulent pipe flow
- LAO+: Focused, Multi-Replica Accelerated Orthogonal Sampling for Robust Binding Free Energy Predictions
- Problem-Specific Basis Quantum State Readout via Proper Orthogonal Decomposition
- Fourier Finite-Difference method and Proper Orthogonal Decomposition for quick spectral study of 2D periodic photonic structures
- Opacity predictions in plasmas under stellar conditions using deep learning
BibTeX
@misc{aksen2026novel,
title = {{A novel machine learning framework based on proper orthogonal decomposition for turbine wake flow predictions in natural waterways}},
author = {Mustafa Meriç Aksen and Hossein Seyedzadeh and Mehrshad Gholami Anjiraki and Ali Khosronejad},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.31547},
eprint = {2610.31547},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31547},
}
RIS
TY - UNPB
TI - A novel machine learning framework based on proper orthogonal decomposition for turbine wake flow predictions in natural waterways
AU - Mustafa Meriç Aksen
AU - Hossein Seyedzadeh
AU - Mehrshad Gholami Anjiraki
AU - Ali Khosronejad
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2610.31547
UR - https://arxiv.org/abs/2610.31547
AB - We developed a machine learning (ML) surrogate trained on instantaneous large eddy simulation (LES) data filtered using proper orthogonal decomposition (POD) to predict mean flow fields in natural rivers with complex geometries. The surrogate enables efficient prediction of mean flow fields downstream of marine hydrokinetic (MHK) turbine arrays at a fraction of the computational cost of high fidelity simulations, facilitating turbine layout optimization in natural waterways where bathymetry, turbulence, and turbine interactions influence the available energy resource. We demonstrate the capabilities of the model at a real world tidal river site under ebb tide conditions. Complex bed bathymetry and planform geometry generate substantial background turbulence that can obscure turbine wake signatures. POD is employed to extract relevant flow structures, improving the ability of the ML model to learn wake dynamics. First, LES is performed at the tidal site to generate training data. POD is then applied to decompose the instantaneous LES fields into spatial modes. Selected modes are used to train a three dimensional convolutional encoder decoder network to predict time averaged velocity fields. This selective filtering reduces the influence of dominant background fluctuations while preserving information relevant to turbine wakes. At equivalent network size and training cost, the POD filtered surrogate reduces the mean squared error by three orders of magnitude compared with the same network trained on unfiltered LES data.
ER -