Wind Farm Modeling with Interpretable Physics-Informed Machine Learning

Howland, Michael F.; Dabiri, John O.

doi:10.3390/en12142716

Cited by 43 publications

(19 citation statements)

References 65 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2) Previous machine learning based wind prediction studies e.g. [16,18,40] treat machine learning models as "black-box" and require the corresponding input and target values for training. Then they can predict the wind patterns which are present in the training dataset.…”

Section: Introductionmentioning

confidence: 99%

Spatiotemporal wind field prediction based on physics-informed deep learning and LIDAR measurements

Zhang

Zhao

2021

Applied Energy

View full text Add to dashboard Cite

Section: Introductionmentioning

confidence: 99%

Spatiotemporal wind field prediction based on physics-informed deep learning and LIDAR measurements

Zhang

Zhao

2021

Applied Energy

View full text Add to dashboard Cite

“…It is worth noting that although our study has revealed some statistical correlations between near-wake behavior and individual turbine parameters, the wake behavior in the field is influenced by multiple intertwining physical processes and cannot be readily predicted with a single turbine parameter. However, it is conceivable that some sophisticated data mining and machine learning approaches (e.g., [76][77][78]) can be employed in the future to yield more accurate and robust predictors of these wake behaviors and can be integrated into the controllers of utility-scale turbines.…”

Section: Conclusion and Discussionmentioning

confidence: 99%

Snow-powered research on utility-scale wind turbine flows

Hong

Abraham

2020

Acta Mech. Sin.

View full text Add to dashboard Cite

This paper provides a review of the general experimental methodology of snow-powered flow visualization and super-large-scale particle imaging velocimetry (SLPIV), the corresponding field deployments and major scientific findings from our work on a 2.5 MW utility-scale wind turbine at the Eolos field station. The field measurements were conducted to investigate the incoming flow in the induction zone and the near-wake flows from different perspectives. It has been shown that these snowpowered measurements can provide sufficient spatiotemporal resolution and fields of view to characterize both qualitatively and quantitatively the incoming flow, all the major coherent structures generated by the turbine (e.g., blade, nacelle and tower vortices, etc.) as well as the development and interaction of these structures in the near wake. Our work has further revealed several interesting behaviors of near-wake flows (e.g., wake contraction, dynamic wake modulation, and meandering and deflection of nacelle wake, etc.), and their connections with constantly-changing inflows and turbine operation, which are uniquely associated with utility-scale turbines. These findings have demonstrated that the near wake flows, though highly complex, can be predicted with substantial statistical confidence using SCADA and structural response information readily available from the current utility-scale turbines. Such knowledge can be potentially incorporated into wake development models and turbine controllers for wind farm optimization in the future.

show abstract

“…Upon the yaw misalignment of an upwind turbine, there is a time lag associated with the advection timescale of the flow for the control decision to influence a downwind turbine. While the advection time depends on the length scale of the turbulent eddy (Del Álamo and Jiménez, 2009;Yang and Howland, 2018;Howland and Yang, 2018), the mean flow advection approximately follows the mean wind speed in wind farms (Taylor, 1938;Lukassen et al, 2018). The number of simulation time steps associated with the approximate advection time between the first and last turbines is computed as…”

Section: Appendix A: Enkf Test Model Problemmentioning

confidence: 99%

Optimal closed-loop wake steering – Part 1: Conventionally neutral atmospheric boundary layer conditions

et al. 2020

Self Cite

View full text Add to dashboard Cite

Abstract. Strategies for wake loss mitigation through the use of dynamic closed-loop wake steering are investigated using large eddy simulations of conventionally neutral atmospheric boundary layer conditions in which the neutral boundary layer is capped by an inversion and a stable free atmosphere. The closed-loop controller synthesized in this study consists of a physics-based lifting line wake model combined with a data-driven ensemble Kalman filter (EnKF) state estimation technique to calibrate the wake model as a function of time in a generalized transient atmospheric flow environment. Computationally efficient gradient ascent yaw misalignment selection along with efficient state estimation enables the dynamic yaw calculation for real-time wind farm control. The wake steering controller is tested in a six-turbine array embedded in a statistically quasi-stationary, conventionally neutral flow with geostrophic forcing and Coriolis effects included. The controller statistically significantly increases power production compared to the baseline, greedy, yaw-aligned control provided that the EnKF estimation is constrained and informed with a physics-based prior belief of the wake model parameters. The influence of the model for the coefficient of power Cp as a function of the yaw misalignment is characterized. Errors in estimation of the power reduction as a function of yaw misalignment are shown to result in yaw steering configurations that underperform the baseline yaw-aligned configuration. Overestimating the power reduction due to yaw misalignment leads to increased power over the greedy operation, while underestimating the power reduction leads to decreased power; therefore, in an application where the influence of yaw misalignment on Cp is unknown, a conservative estimate should be taken. The EnKF-augmented wake model predicts the power production in yaw misalignment with a mean absolute error over the turbines in the farm of 0.02P1, with P1 as the power of the leading turbine at the farm. A standard wake model with wake spreading based on an empirical turbulence intensity relationship leads to a mean absolute error of 0.11P1, demonstrating that state estimation improves the predictive capabilities of simplified wake models.

show abstract

Wind Farm Modeling with Interpretable Physics-Informed Machine Learning

Cited by 43 publications

References 65 publications

Spatiotemporal wind field prediction based on physics-informed deep learning and LIDAR measurements

Spatiotemporal wind field prediction based on physics-informed deep learning and LIDAR measurements

Snow-powered research on utility-scale wind turbine flows

Optimal closed-loop wake steering – Part 1: Conventionally neutral atmospheric boundary layer conditions

Contact Info

Product

Resources

About