Wind Turbine Load Prediction Using Quasi-Random Simulation Grids
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining loads on wind turbines are inefficient due to the complexity of simulating all possible parameter combinations, which is resource-intensive and time-consuming, making it difficult to assess site-specific mechanical load-bearing capacity and predict loads accurately.
Innovation Solution
A method using a grid structure with quasi-randomly distributed grid points for wind conditions, combined with load simulations and machine learning models, allows for predicting loads on wind turbines by training on a subset of parameter combinations, reducing the need for extensive simulations and enabling faster, more accurate load predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all possible parameter combinations are simulated to ensure accurate load predictions, then prediction accuracy is improved, but computational resources and time increase significantly
Solution Approach 1:
The method pre-simulates a representative subset of parameter combinations offline to build a training dataset, which is then used to train a prediction model. This preliminary action allows the model to make rapid predictions without requiring time-consuming full simulations at runtime, thus resolving the contradiction between accuracy and simulation time.
Solution Approach 2:
Instead of performing actual simulations for every possible parameter combination, the method creates a predictive model that copies the behavior patterns learned from a limited set of simulations. The model acts as a surrogate that replicates simulation results for any input parameters, eliminating the need for exhaustive simulations while maintaining prediction accuracy.
2Adaptability or versatility
If all possible parameter combinations are simulated to cover the complete parameter space, then comprehensive load assessment is improved, but computational resources and memory usage increase significantly
Solution Approach 1:
The method transforms the problem from directly simulating all parameter combinations to changing the approach by training a machine learning model on a subset of simulations. The model then handles parameter variations through learned patterns rather than explicit simulation, reducing computational complexity while maintaining comprehensive parameter space coverage capability.
Solution Approach 2:
A machine learning prediction model is introduced as an intermediary between the limited simulation data and the requirement for comprehensive parameter space coverage. This intermediary learns from the simulated data and generalizes to unseen parameter combinations, enabling broad coverage without direct simulation of all cases.
3Productivity
If a lean design with minimal load reserves is used to reduce costs, then manufacturing efficiency is improved, but the risk of mechanical overload increases
Solution Approach 1:
The method provides detailed, site-specific load predictions that feed back into the design and operational decisions. By accurately predicting actual loads at specific locations, the system enables manufacturers to optimize designs with minimal load reserves while ensuring reliability through precise load knowledge, rather than using conservative universal safety factors.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to methods for determining loads on wind turbines, comprising creating a grid structure with grid points from randomly varied wind conditions as parameters, in particular comprising turbulence, turbulence intensity and/or wind shear, load simulation of a wind turbine for each of the grid points, providing a prediction model for predicting loads based on wind conditions, training the prediction model using the load simulations performed for each of the grid points, and determining the loads for any combination of wind conditions using the prediction model. The invention also relates to a corresponding training dataset.