Presentations

Abstract

CFD.ML is a fast Machine Learning surrogate model for DNV’s high-fidelity RANS CFD (DNV CFD) simulations predicting turbine interaction losses (blockage and wakes). CFD.ML 2.6 was recently selected as DNV’s default turbine interaction model for offshore Energy Production Assessments /1/. However, with less training data from onshore sites, this model's application was limited to offshore projects. We now share work detailing quality tests to support application of a new model to onshore wind farm sites, with appropriate guardrails. The Problem:  Compared to offshore wind farms, turbine interaction modelling at onshore wind farms must consider: * The influence of terrain on wakes and blockage /2/ * Highly irregular layouts * More variable and diurnal atmospheric conditions (due to land surface heating/cooling) * Generally higher turbulence and higher shear. Terrain effects are only included implicitly by prescribing a non-homogeneous background flow field. CFD.ML explicitly accounts for points 2-4, using 8 free inputs parameters to describe atmospheric conditions.  For CFD.ML 2.6, and earlier, the range of layouts and turbine technologies in the model training data was extensive. However, training data covering onshore site-specific atmospheric conditions was very limited, relative to offshore sites. Significant holes in this parameter space for onshore sites means the model risks hallucination. Methodology:   We propose a model quality test taking turbine interaction modelling results from across a sample of more than 100 onshore wind farms previously assessed by DNV as predicted by CFD.ML 2.6, CFD.ML 2.8, and DNV’s established onshore turbine interaction model. We’ll illustrate some outlier cases exhibiting possible “hallucination” behaviours in the CFD.ML 2.6 results, then show how such outliers are reduced in CFD.ML 2.8 with extra training and by using guardrails restricting the possible choices of atmospheric conditions inputs. We hypothesize that the 8 free parameters aren’t truly independent, and that we can discretise their variations and greatly limit their combinations, whilst still making useful turbine interaction loss predictions. To define the guardrails, we define a small set of discrete atmospheric conditions classes well represented in the ML model training, and derive frequency of those classes using a method based on ERA5 data explored in /3/. For a small set of onshore projects, we also validate predictions from CFD.ML 2.8 against DNV CFD results (those sites being excluded from CFD.ML 2.8 training). At these sites results using site-specific atmospheric conditions, from the DNV CFD analysis, are compared to the results using inputs with guardrails. We will discuss the appropriateness of the guardrails, and the trade-off of site-specificity against the global model robustness. Based on the outcomes here, DNV will comment on the readiness of CFD.ML 2.8 for application to onshore turbine interaction modelling, and where it is expected to outperform existing approaches. References: /1/            “Offshore turbine interaction modelling using CFD.ML”, DNV white paper, 2026 /2/            Bleeg, J.: A numerical study of the influence of terrain on wakes, blockage, wind farm efficiency, and turbine efficiency, Wind Energ. Sci. Discuss. [preprint], https://doi.org/10.5194/wes-2025-291, in review, 2026. /3/            Fernandes, M et.al: “Towards a Global Map of Atmospheric Boundary Layer Properties”, WindEurope, Madrid, 2026