Frequency-Decomposed Deep Koopman Network for Long-Horizon Wind Turbine State Prediction

Abstract Wind turbine dynamics exhibit inherent multiscale characteristics, wherein low-frequency operational trends driven by pitch and torque control coexist with high-frequency stochastic fluctuations arising from atmospheric turbulence, blade–tower interactions, and drive-train torsional oscillations. Conventional single-scale deep learning models cannot simultaneously satisfy the opposed prediction demands that these two regimes impose and tend to either sacrifice long-term stability or fail to represent nonlinear transient behavior. This paper proposes the frequency-decomposed deep Koopman network (FD-DKN), a physics-informed hybrid architecture that overcomes this difficulty by combining fast Fourier transform–based spectral decomposition with two physically motivated prediction branches. A stability-constrained Koopman linear modeling module with spectral-radius constraints guarantees bounded multistep prediction errors for low-frequency dynamics, while a Transformer–bidirectional long short-term memory (BiLSTM) module handles residual high-frequency stochastic components. The decomposition employs adaptive energy thresholding aligned with turbine modal frequencies, and physical control inputs (wind speed, pitch and torque setpoints) are embedded directly into the latent linear dynamics. Evaluated on offshore supervisory control and data acquisition data from a wind farm in Jieyang, China (86,399 samples at 1-s resolution), FD-DKN reaches a spectral coherence of 0.9800, Wasserstein distance of 0.0980, and concordance correlation coefficient of 0.9363 at a 500-step prediction horizon, outperforming five baselines including Seq2Seq-LSTM, temporal convolutional networks, convolutional neural network–Transformer, Transformer–BiLSTM, and sparse dynamic mode decomposition. With 152,244 parameters and a 0.10 ms per-step inference latency, the model meets the real-time deployment requirements of wind turbine control systems operating at 10–100 Hz. Ablation analysis confirms that frequency decomposition, Koopman linear modeling, and control-input embedding are each individually necessary components of the architecture.

Authors

Institutions

Publication Details

Journal
Journal of Energy Engineering
Published
2026-11-13
DOI
https://doi.org/10.1061/jleed9.eyeng-7205
Primary Topic
Model Reduction and Neural Networks
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Frequency-Decomposed Deep Koopman Network for Long-Horizon Wind Turbine State Prediction

Haonan Cui, Xiaodong Wang
5 links
Journal of Energy Engineering
Model Reduction and Neural Networks
article

Frequency-Decomposed Deep Koopman Network for Long-Horizon Wind Turbine State Prediction

Haonan Cui, Xiaodong Wang
article en

Abstract

Abstract Wind turbine dynamics exhibit inherent multiscale characteristics, wherein low-frequency operational trends driven by pitch and torque control coexist with high-frequency stochastic fluctuations arising from atmospheric turbulence, blade–tower interactions, and drive-train torsional oscillations. Conventional single-scale deep learning models cannot simultaneously satisfy the opposed prediction demands that these two regimes impose and tend to either sacrifice long-term stability or fail to represent nonlinear transient behavior. This paper proposes the frequency-decomposed deep Koopman network (FD-DKN), a physics-informed hybrid architecture that overcomes this difficulty by combining fast Fourier transform–based spectral decomposition with two physically motivated prediction branches. A stability-constrained Koopman linear modeling module with spectral-radius constraints guarantees bounded multistep prediction errors for low-frequency dynamics, while a Transformer–bidirectional long short-term memory (BiLSTM) module handles residual high-frequency stochastic components. The decomposition employs adaptive energy thresholding aligned with turbine modal frequencies, and physical control inputs (wind speed, pitch and torque setpoints) are embedded directly into the latent linear dynamics. Evaluated on offshore supervisory control and data acquisition data from a wind farm in Jieyang, China (86,399 samples at 1-s resolution), FD-DKN reaches a spectral coherence of 0.9800, Wasserstein distance of 0.0980, and concordance correlation coefficient of 0.9363 at a 500-step prediction horizon, outperforming five baselines including Seq2Seq-LSTM, temporal convolutional networks, convolutional neural network–Transformer, Transformer–BiLSTM, and sparse dynamic mode decomposition. With 152,244 parameters and a 0.10 ms per-step inference latency, the model meets the real-time deployment requirements of wind turbine control systems operating at 10–100 Hz. Ablation analysis confirms that frequency decomposition, Koopman linear modeling, and control-input embedding are each individually necessary components of the architecture.

Journal of Energy EngineeringVol. 153(1)
Shenyang University of Technology (CN), Shenyang Institute of Science and Technology (CN)
Affordable and clean energy, Climate action
Model Reduction and Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon