Research on the Derated Power Data Identification Method of a Wind Turbine Based on a Multi-Gaussian-Discrete Joint Probability Model

Yuanchi Ma, Yongqian Liu, Zhiling Yang, Jie Yan, Tao Tao, David Infield2022 · DOI

TL;DR

This paper introduces a novel method for identifying normal, derated, and abnormal operational data from wind turbines using a multi-Gaussian-discrete joint probability model. By treating the derated power state as a latent variable and employing the Expectation-Maximization (EM) algorithm, the method effectively distinguishes between different operating conditions, crucial for accurate performance monitoring and fault diagnosis.

What Problem Does This Paper Address?

Wind turbines generate vast amounts of operational data through SCADA systems, which is vital for performance evaluation, prediction, and grid dispatch. However, this data often contains noise, measurement errors, and importantly, periods where the turbine operates in a "derated" state (intentionally producing less power than its maximum capacity). Existing methods often treat this derated data as outliers, leading to a loss of valuable information. Furthermore, there's no direct indicator in SCADA data for derated operation, making it challenging to identify and analyze. This paper addresses the need for a method that can accurately identify and classify normal, various levels of derated power, and truly abnormal data, rather than simply discarding potentially useful derated operational data.

Methods

The core of the proposed method is a multi-Gaussian-discrete joint probability distribution model. This model characterizes the relationship between wind speed (X) and power output (Y) by incorporating a latent variable (Z) representing the derated power state. This latent variable can take on K different values, corresponding to the normal operating state and K-1 distinct derated power levels.

The model makes several key assumptions:

  • Power-limited state assumption: The derated power state can be categorized into a finite number of levels.
  • Power-limited state output assumption: The actual power output in a derated state is the theoretical power curve multiplied by a derated power coefficient ($\alpha_k$), where $\alpha_k=1$ for normal operation and $\alpha_k < 1$ for derated states.
  • Wind speed discretization: Wind speed is divided into J equal intervals, and each interval is represented by its median.
  • Gaussian distribution for power: Given a wind speed and derated state, the power output follows a Gaussian distribution.

To estimate the parameters of this complex model, the authors employ the Expectation-Maximization (EM) algorithm. This iterative algorithm is well-suited for models with latent variables. It involves:

  • E-step: Calculating the posterior probability of the derated state for each data point.
  • M-step: Updating the model parameters (derated power coefficients, variances, wind speed distribution parameters) to maximize the likelihood function.

The paper also introduces a method for optimal derated power level identification using a k-means-inspired approach to determine the number of derated states (K) and for parameter initialization to ensure stable convergence of the EM algorithm.

Datasets

The proposed method was validated using SCADA data from three 2.5 MW direct-drive permanent-magnet synchronous generator wind turbines (#1, #2, and #6) located in North China. The data was collected over a period from November 20, 2014, to January 19, 2015. The paper does not introduce new benchmark datasets but uses real-world operational data representative of a region with common wind curtailment practices.

Key Insights and Ideas

  • Derated power is not just noise: The paper highlights that derated power operation is a significant and informative aspect of wind turbine operation, not merely an anomaly to be discarded.
  • Latent variable modeling for states: Effectively modeling distinct operating states (normal, derated levels) by treating them as unobserved (latent) variables is a powerful approach.
  • EM algorithm's suitability: The EM algorithm is a robust choice for parameter estimation in probabilistic models with latent variables, enabling the extraction of complex relationships from data.
  • Universality of the approach: The method demonstrates a degree of universality, being effective across turbines with different derated power characteristics and distributions.
  • Threshold flexibility: The choice of threshold for identifying outliers based on posterior probability offers flexibility, allowing users to balance data preservation with outlier removal based on specific quality requirements.

Key Results

  • The proposed method successfully identified and classified normal operation and multiple levels of derated power operation for the tested wind turbines.
  • For turbine #2, the method identified three operating states (normal and two derated levels), consistent with visual inspection of the data.
  • For turbine #1, the method identified five operating states, including normal and four derated power levels, demonstrating its ability to handle more complex scenarios.
  • The method effectively separated derated power data from normal operating data, providing a clearer picture of turbine performance.
  • The outlier detection was shown to be reliable, even when dealing with different numbers of derated power levels.
  • The impact of the outlier detection threshold was investigated, showing that while it influences the number of retained data points, the core identification of meaningful operational data remains robust.

Limitations and Future Work

  • Local Optima: The EM algorithm, while effective, can converge to local optima, meaning the final model parameters are sensitive to the initial values. The paper addresses this by proposing a parameter initialization method but acknowledges that achieving a global optimum is not guaranteed.
  • Parameter Selection: While the method is shown to be robust to threshold selection, other parameters within the model might require careful tuning for optimal performance in diverse scenarios.
  • Future Work: The paper implicitly suggests that further refinement of the parameter initialization and exploration of more advanced clustering techniques could enhance the model's performance. Additionally, investigating the impact of different wind turbine types and environmental conditions on the model's effectiveness would be valuable.

Summary updated

BibTeX
@misc{ma2022research,
  author = {Yuanchi Ma and Yongqian Liu and Zhiling Yang and Jie Yan and Tao Tao and David Infield},
  title = {{Research on the Derated Power Data Identification Method of a Wind Turbine Based on a Multi-Gaussian-Discrete Joint Probability Model}},
  year = {2022},
  doi = {10.3390/s22228891},
}
Figure