Comparison of Data-Driven Modeling Approaches for Control Optimization of Floating Offshore Wind Turbines

TL;DR

This work tackles the challenge of expensive high-fidelity simulations for design optimization of floating offshore wind turbines by developing a derivative-function surrogate model (DFSM) built on a linear parameter-varying (LPV) state-space framework. The DFSM is trained from time-series data to approximate state derivatives and outputs, enabling accurate closed-loop simulations and design-of-experiments far faster than OpenFAST. Comparisons with subspace identification (n4sid) and LSTM show that the DFSM offers a favorable balance of computational speed and predictive accuracy, achieving roughly a 48× speedup while maintaining key dynamic features such as blade-pitch and power signals, though some nonlinear outputs like tower-base moment are more challenging. The DFSM is demonstrated in two use cases—closed-loop control simulations and DOE-driven DEL analysis—highlighting its practical potential for rapid controller tuning and early-stage design exploration, with avenues for extending to more nonlinear models and broader state sets.

Abstract

Models that balance accuracy against computational costs are advantageous when designing wind turbines with optimization studies, as several hundred predictive function evaluations might be necessary to identify the optimal solution. We explore different approaches to construct low-fidelity models that can be used to approximate dynamic quantities and be used as surrogates for design optimization studies and other use cases. In particular, low-fidelity modeling approaches using classical systems identification and deep learning approaches are considered against derivative function surrogate models ({DFSMs}), or approximate models of the state derivative function. This work proposes a novel method that utilizes a linear parameter varying (LPV) modeling scheme to construct the DFSM. We compare the trade-offs between these different models and explore the efficacy of the proposed DFSM approach in approximating wind turbine performance and design optimization studies for controllers. Results show that the proposed DFSM approach balances computational time and model accuracy better than the system identification and deep learning based models. Additionally, the DFSM provides nearly a fifty times speed-up compared to the high-fidelity model, while balancing accuracy.

Figures (15)

• Figure 1: A high-level workflow diagram illustrating the different components associated with simulating and/or performing design optimization studies for a wind turbine system.
• Figure 2: Generator torque and blade pitch vs. wind speed.
• Figure 3: Two different surrogate modeling approaches that can be used for wind turbines. The first approach shown in Fig. \ref{['fig:surrogate1']} can be used to train a surrogate model that predicts key performance metrics given information about the load cases. The second approach, shown in Fig. \ref{['fig:surrogate2']}, has been used extensively to construct surrogate models that can be used for design optimization studies.
• Figure 4: An alternate approach to construct a surrogate model that can be used for predicting time-series of key outputs, given inputs and controls.
• Figure 5: Wind speed trajectory and histogram plots for three different load cases with $\bar{\bm{w}} = [12,14,16]$ m/s.