TL;DR
This paper introduces the Bayesian Ascent (BA) method, a novel data-driven optimization algorithm that efficiently maximizes wind farm power by coordinating individual turbine controls. BA combines Bayesian Optimization with a trust region strategy to achieve rapid, monotonic power increases with minimal data, even in noisy experimental settings.
What Problem Does This Paper Address?
Modern wind farms face a challenge: individual turbines are often optimized for maximum power output without considering their impact on neighboring turbines. This "greedy" approach leads to wake effects – where the airflow from one turbine disrupts the wind for those downstream, reducing overall farm efficiency. While cooperative control strategies aim to mitigate this by coordinating turbine actions (like adjusting blade pitch or yaw angle), these methods often rely on complex, analytical models of wind farm power that may not accurately reflect real-world conditions. Furthermore, many existing data-driven approaches require a large amount of data to find optimal settings, making them impractical for real-time control. This paper addresses the need for a data-efficient, model-free optimization method that can find optimal cooperative control strategies for wind farms in real-time, even with limited and noisy data.
Methods
The core contribution is the Bayesian Ascent (BA) method. BA builds upon Bayesian Optimization (BO), a technique that uses a probabilistic model (specifically, Gaussian Process regression) to learn an unknown objective function from a small number of data points. BO iteratively selects the next point to sample by balancing exploration (reducing uncertainty) and exploitation (seeking higher objective values).
What's novel about BA is its integration of a trust region constraint into the BO framework. Instead of sampling from the entire input space, BA restricts its search to a hypercube region around the current best-known solution. This constraint, inspired by trust region optimization algorithms, ensures that the algorithm attempts to find solutions that monotonically increase the target value (wind farm power). The size of this trust region is adaptively adjusted: it expands when significant improvements are observed, accelerating convergence, and resets to an initial size if progress stalls, preventing the algorithm from getting stuck in suboptimal areas. This combination allows BA to efficiently explore the parameter space and converge towards an optimum with fewer samples than traditional BO methods.
Datasets
The paper uses two main sources of data:
- Analytical Functions: For initial validation, the BA method is tested on a 2D quadratic function to demonstrate its ability to find optima and illustrate its convergence behavior.
- Experimental Data: The primary evaluation is conducted using a laboratory setting with two scaled wind turbine models. Data is collected from these physical turbines, including measurements of power output, pitch angle, yaw angle, and load resistance. This experimental setup is used to test both single-turbine optimization and cooperative control of two turbines.
Key Insights and Ideas
- Wake Interference is a Major Bottleneck: The paper clearly illustrates how the greedy control of individual turbines, while maximizing their own output, can significantly decrease the power output of downstream turbines due to wake effects.
- Cooperative Control Offers Significant Gains: By coordinating control actions, the overall wind farm power can be substantially increased, even if it means slightly reducing the power of an upstream turbine.
- Data Efficiency is Crucial for Real-World Application: Traditional optimization methods often require extensive data. BA's ability to learn from a small number of noisy measurements is a key enabler for practical, real-time control of physical systems like wind farms.
- Trust Region Strategy Enhances Monotonicity and Convergence: The integration of a trust region constraint is critical. It guides the search towards solutions that are likely to improve the objective function, leading to more stable and predictable power increases and faster convergence compared to standard BO.
- Probabilistic Modeling Handles Noise Effectively: Gaussian Process regression, a core component of BA, is adept at modeling complex relationships and providing uncertainty estimates, which is vital for dealing with the inherent noise in wind turbine power measurements.
Key Results
- Analytical Function Optimization: BA demonstrated rapid convergence and an almost monotone increase in the target value when optimizing a 2D quadratic function.
- Single Wind Turbine Optimization: In a laboratory setting, BA successfully identified near-optimal pitch angles and load resistances to maximize the power output of a single scaled wind turbine, achieving results close to those found through exhaustive testing.
- Cooperative Wind Turbine Optimization: When applied to two scaled wind turbines, BA was able to find coordinated control actions for the upstream turbine that significantly increased the total farm power. This was achieved by adjusting the upstream turbine's yaw and pitch angles to mitigate wake interference, even with noisy power measurements. The results showed that BA could increase total power by executing a small number of trial control actions.
Limitations and Future Work
- Local Optima: The authors acknowledge that, similar to trust region methods, BA is likely to find a local optimum rather than a guaranteed global optimum, especially when starting from a suboptimal initial point.
- Computational Cost: Building the Gaussian Process model can be computationally expensive, with complexity scaling cubically with the number of data points. However, the authors note that BA is designed for scenarios with a small number of data points (less than a hundred), mitigating this concern.
- Noise Sensitivity: While BA handles noise well, the paper mentions that the relatively large variance in power measurements (due to wind fluctuations) can make it challenging for the algorithm to precisely follow the ascending direction of the power curve.
- Future Work: The paper suggests that BA's ability to efficiently find local optima quickly makes it suitable for real-time control applications. Further exploration of strategies to improve global optimization or to better handle extreme noise levels could be beneficial. The authors also highlight the potential for BA to be applied to more complex wind farm configurations with a larger number of turbines.
Summary updated