TL;DR
This paper introduces a data-driven, cooperative control strategy using the Bayesian Ascent (BA) algorithm to maximize wind farm power output by coordinating individual turbine actions. Wind tunnel experiments with scaled turbines demonstrate that this cooperative approach significantly boosts overall power generation compared to traditional greedy control, especially in dense wind farm configurations.
What Problem Does This Paper Address?
Modern wind farms face a significant challenge: the "wake effect." When wind flows through a turbine, it creates a turbulent wake of reduced wind speed and increased turbulence behind it. Downstream turbines in this wake produce less power. Conventional control strategies operate each turbine independently to maximize its own output (a "greedy" approach), which exacerbates the wake effect and leads to suboptimal total farm power. This paper addresses the need for a cooperative control strategy that accounts for these interactions to maximize the entire wind farm's power output.
Methods
The core of this work is the Bayesian Ascent (BA) algorithm, a data-driven optimization method designed for efficiently finding optimal control actions with minimal data. BA combines:
- Gaussian Process (GP) Regression: This probabilistic model learns the relationship between control actions (yaw angle, blade pitch) and wind farm power output. It provides not only a prediction of power but also an estimate of uncertainty, crucial for guiding the search.
- Trust Region Optimization: This technique constrains the search for new control actions to a "trust region" around the current best-performing actions. This ensures that the algorithm makes monotonic progress towards higher power output, preventing erratic steps and improving convergence.
The BA algorithm iteratively performs three steps:
- Learning: Updates the GP model with new data.
- Optimization: Selects the next set of control actions to try, aiming to maximize expected improvement.
- Observation: Executes the chosen actions, measures the resulting power output, and adds this to the dataset.
This process is applied to a cooperative control problem where all turbines' actions are optimized simultaneously to maximize total farm power, contrasting with the decentralized greedy approach.
Datasets
The paper does not use pre-existing datasets for training or evaluation. Instead, it relies on wind tunnel experiments using six scaled wind turbine models. This setup generates the necessary input-output data (control actions and measured power) for the BA algorithm to learn and optimize. The experiments are conducted under various configurations (number of turbines, wind direction, farm layout) to simulate different real-world scenarios.
Key Insights and Ideas
- Wake Interference is a Major Bottleneck: The experiments clearly show that downstream turbines suffer significantly from the wake of upstream ones, highlighting the inefficiency of independent control.
- Cooperation Pays Off: By coordinating turbine actions, particularly by slightly misaligning upstream turbines (yaw offset), the wake can be steered away from downstream turbines. This leads to a more even distribution of power and a substantial increase in total farm output.
- BA is Efficient and Robust: The BA algorithm demonstrates a remarkable ability to find improved control strategies with a relatively small number of trial actions. Its probabilistic nature and trust region constraint allow it to navigate the complex, high-dimensional optimization problem effectively.
- Scalability Matters: As the number of turbines increases, the wake effect becomes more pronounced, and the relative improvement from cooperative control also increases, suggesting its critical importance in larger wind farms.
- Wind Direction Influences Strategy: The optimal cooperative control strategy, particularly the yaw offset, is sensitive to wind direction. Asymmetric configurations and wind direction deviations require careful adjustments to effectively manage wake interactions.
Key Results
- Significant Power Gains: The cooperative control strategy, implemented via the BA algorithm, consistently improved total wind farm power production. For instance, in Scenario 1, the relative gain in power efficiency ranged from approximately 24.8% to 27.3% with increasing numbers of turbines. Scenario 2 showed gains up to 35.7% for certain wind directions.
- Monotonic Improvement: The BA algorithm generally showed monotonic increases in wind farm power efficiency throughout the experiments, a testament to its trust region constraint.
- Reduced Iterations: Compared to other optimization methods (like Genetic Algorithms mentioned in the introduction), BA requires significantly fewer iterations to reach a near-optimal solution.
- Effective for Different Layouts: The cooperative strategy proved effective for both linear and grid-like wind farm configurations, indicating its broad applicability.
Limitations and Future Work
- Idealized Conditions: The wind tunnel experiments were conducted under constant wind speed and direction. In a real wind farm, these conditions are dynamic, which poses a significant challenge for real-time control.
- Convergence Time: While BA is efficient, the convergence time increases with the number of turbines due to the higher dimensionality of the control problem. The time for wind flow to stabilize after control adjustments also adds to the overall time.
- True Optimum Unknown: The paper acknowledges that the absolute optimal cooperative control strategy is generally unknown. BA aims to improve upon the initial strategy rather than guarantee finding the global optimum.
Future work suggested includes:
- Developing strategies for handling varying wind speeds and directions, potentially by running multiple BA instances or incorporating wind conditions as part of the GP model.
- Investigating experimental design to test and validate control strategies more comprehensively.
- Exploring parallel BA algorithms for faster convergence in larger wind farms.
Summary updated