Wind Turbine Wake Control via Machine Learning Surrogate Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current wind farm control methods fail to effectively address short-term wake interference between turbines, leading to reduced power production and increased mechanical loads, as they rely on simplified models that do not account for dynamic conditions and require high computational costs.
Innovation Solution
A data-driven approach using machine learning to model dependencies between upstream and downstream turbines, predicting power production ratios and determining control parameters to mitigate wake effects without physical assumptions or numeric simulations, allowing for dynamic adjustments to optimize energy production and reduce turbine fatigue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If simplified wake models (e.g., Jensen wake model) are used to construct analytical wind farm power functions, then computational cost is reduced, but the models do not accurately reflect the conditions of a wind farm site or wind turbine model
Solution Approach 1:
The patent uses a data-driven approach that creates a virtual copy of the complex CFD simulation results through machine learning models. Instead of running expensive CFD simulations in real-time, the system trains on CFD data to create surrogate models that replicate wake behavior with much lower computational cost while maintaining high accuracy.
Solution Approach 2:
The patent replaces traditional analytical wake models (Jensen, etc.) with data-driven machine learning models. These ML models substitute the simplified physical assumptions of analytical models with patterns learned from high-fidelity CFD simulations, achieving both computational efficiency and accuracy.
2Measurement precision
If high-fidelity Computational Fluid Dynamics (CFD) simulation is used to construct the parametric wind farm power function, then accuracy is improved, but computational cost increases significantly
Solution Approach 1:
The patent performs CFD simulations in advance during an offline training phase to generate comprehensive datasets. These pre-computed results are then used to train machine learning models, so that during online operation, only the lightweight ML inference is needed, achieving real-time performance with pre-performed heavy computation.
Solution Approach 2:
The patent creates surrogate models that copy the input-output behavior of expensive CFD simulations. These surrogate models are trained on CFD data and then used to predict wake effects without running actual CFD simulations, maintaining accuracy while reducing computational cost by orders of magnitude.
3Device complexity
If traditional wake models provide only time-independent solutions, then computational complexity is reduced, but it is not possible to use the models to derive short-term adjustments of turbine controllers which can dynamically react to wake effects
Solution Approach 1:
The patent transitions from static, time-independent wake models to dynamic models that capture temporal variations in wake behavior. The machine learning models are trained on time-series CFD data that includes evolving wake structures, enabling the system to predict and respond to dynamic wake effects in real-time.
Solution Approach 2:
The patent implements a closed-loop control system where wake predictions from the data-driven model feed back to adjust turbine control parameters dynamically. This feedback mechanism enables short-term adjustments of turbine controllers to respond to changing wake conditions, improving both power production and wake mitigation.
Data Source
AI summary
Provided is an apparatus and method for cooperative controlling wind turbines of a wind farm, wherein the wind farm includes at least one pair of turbines aligned along a common axis approximately parallel to a current wind direction and having an upstream turbine and a downstream turbine. The method includes the steps of: a) providing a data driven model trained with a machine learning method and stored in a database, b) determining a decision parameter for controlling at least one of the upstream turbine and the downstream turbine by feeding the data driven model with the current power production of the upstream turbine which returns a prediction value indicating whether the downstream turbine will be affected by wake, and/or the temporal evolvement of the current power production of the upstream turbine; c) based on the decision parameter, determining control parameters for the upstream turbine and/or the downstream turbine.