Gaussian processes modeling for floating offshore wind turbine fatigue in wind farm context based on input/output dimension reduction

To support the development of electricity production from wind power, IFP Energies nouvelles is involved in the energy transition as a research and training player, especially in emerging technologies such as floating offshore wind turbines (FOWT). 
The main objective of the design of a wind turbine platform is to obtain a robust solution to environmental conditions and at a lower construction cost. This means that the wind turbine must be able to withstand a set of extreme scenarios but also that its fatigue damage must be controlled over time.
The fatigue design of an offshore wind turbine is an expensive task because it requires a large volume of multi-physics simulations in order to cover a large number of environmental conditions (wind and sea), especially if the wind turbine is placed within a farm (wind wake effect). In practice during an industrial study, the limitations due to the computation time force to limit the input parameters to a few scalars. This reduces the robustness of the final design and increases its uncertainty. 
The need for an accurate non-intrusive surrogate model to accelerate computations is becoming apparent. A surrogate model designates a function built as an approximation of a numerical simulator. This approximation is constructed from the simulator output values at different points in the input space. A classical technique (in the field of uncertainty quantification and in machine learning) for building a metamodel consists in modeling a simulator using a Gaussian process and obtaining an approximation by calculating the a posteriori mean of the process. 
An efficient surrogate model could especially open the gate to more comprehensive uncertainty quantification (UQ) studies and design optimization. This will also allow real-time control, live tracking of wind turbines fatigue, and thus a more efficient maintenance planning.
The state of the art regarding surrogate modelling for WT computations mainly address low dimensional input spaces (e.g., wind or wave statistical parameters) and low dimensional output spaces (e.g., DEL  at a particular point of the structure). Hence, the main contribution of the thesis will be to develop a non-intrusive surrogate modelling strategy that would have the ability to handle both input and output space high dimensions, typically O(104-106), for WT computations. More precisely, the surrogate strategy should ultimately tackle wind/wave time series input to predict space variant fields.
We suggest splitting the thesis work into 4 steps with increasing complexity to ensure the progressivity of the subject and thus secure its handling and its implementation by the future candidate:
1.    Production of a state of the art on the metamodeling techniques for computer simulations and more specifically on the metamodeling techniques used in the wind energy context.
2.    Development of a non-intrusive low-dimensional metamodeling strategy in a farm context (wind wake) for the prediction of AEP and DEL at a given point of the structure. Some reflexions could be made on the incident wind field parameterization. This reflexion could also draw on the basis of the work already carried out in the department regarding the wake effect in wind farm.
3.    Production of a state of the art on metamodeling techniques combined with input/output dimension reduction.
4.    Development of:
a.     a non-intrusive high-dimensional output metamodeling strategy for the reconstruction of a time variant variable or the prediction of a space variant field,
b.    Development of a non-intrusive high-dimensional input and output metamodeling strategy for the reconstruction of a time variant variable or the prediction of a space variant field with wind transient and possibly wave transient as inputs.

Keywords: Data science, offshore wind energy, gaussian processes, mechanical design, machine learning

• Academic supervisor    Emmanuel Vazquez, professeur, The Laboratory of Signals and System (L2S), CentraleSupélec, Paris-Saclay Univ.
• Doctoral School    Doctoral school 422 STITS, http://ed-stits.fr/fr/
• IFPEN supervisor     Nicolas Bonfils, research engineer, IFP Énergies Nouvelles (IFPEN)
• PhD location    IFPEN/L2S
• Duration and start date    3 years, beginning of the fourth trimester
• Employer    IFPEN, Rueil-Malmaison, France
• Academic requirements    Engineering degree with a specialization in data science or statistics
• Language requirements    Very good knowledge of French is required, knowledge of English is desirable
• Other requirements     Knowledge of python, R and main machine learning libraries is required. Knowledge of GitHub repository service is desirable