DNV has combined high-fidelity computational fluid dynamics (CFD) models with machine learning (ML) to create a next-generation prediction tool for use in the design and analysis of wind farms. Thus far, results look promising, and the model is already being used to supplement CFD analyses, improving accuracy and reducing the number of CFD simulations required.

In 2017 and 2018, DNV presented and published research indicating that there is a bias in flow models used to predict the energy production of a prospective wind farm. These models assumed that wind turbines could only affect one another through their wakes, but the research clearly demonstrated that blockage is also a significant contributor to turbine interaction loss. 

The wind industry now recognizes that a new approach to turbine interaction analysis is needed, but at present lacks an option that can efficiently and reliably account for both wakes and blockage when modelling wind farm flows. Traditional wind farm flow models are fast but ignore blockage effects; high-fidelity flow models are more complete and accurate, but turnaround times can be relatively long. 

A surrogate model for the prediction of turbine interaction loss 

The combination of CFD and machine learning offers a solution. When trained on results from many CFD-based wind farm simulations, an appropriately designed machine learning model can encode the turbine interaction patterns evident in the CFD results. The model, which we call CFD.ML, can then use these encoded patterns to accurately predict CFD results for wind farm configurations not seen before.

The interaction between an array of turbines and the atmosphere is complex, involving a range of atmospheric phenomena and scales. CFD is capable of capturing the main physical drivers behind turbine interaction loss, but it requires the brute force of a supercomputer. CFD.ML, however, can deliver a similar level of accuracy at much greater speed using just a desktop computer—all while side-stepping many of the simplifying assumptions used in typical flow models. The speed increase is substantial: with just 2% of the computational resource, CFD.ML can deliver results 2 million times faster than the CFD model upon which it was trained.

The benefits 

CFD and CFD.ML form a symbiotic pair. With limited time and computational resource, CFD cannot practically cover the full range of wind farm operation; CFD.ML can quickly fill in the gaps between CFD simulations, increasing accuracy and reducing cost. In turn, training on an ever-growing set of CFD results progressively improves accuracy and range of applicability of the CFD.ML model. 

CFD.ML, once trained, could also potentially be used without CFD. DNV’s WindFarmer Analyst software team is currently evaluating this possibility. A successful evaluation would signal a significant step forward in the design and analysis of wind farms, which has typically been the domain of simplified models. Unlike these models, CFD.ML accounts for wakes and blockage together while also representing the impact of important influences, such as atmospheric stability. 

CFD.ML is a timely development and DNV believes there is significant market potential with many companies in the wind segment looking for a better, faster model of wind farm flows.