Power Prediction from Wind Turbine SCADA Data in the Presence of Modelling and Measurement Uncertainties

Abstract

Wind turbines are continuously exposed to harsh environmental and operational conditions throughout their lifetime, leading to the gradual degradation of their components. If left unaddressed, these degraded components can adversely affect turbine performance and significantly increase the likelihood of failure. As degradation progresses, the risk of failure escalates, making it essential to implement appropriate risk control measures.

One effective risk control method involves performing inspection and monitoring activities that provide valuable insights into the condition of the structure, enabling the formulation of appropriate maintenance strategies based on accurate assessments.

Supervisory Control and Data Acquisition (SCADA) systems offer low-resolution condition-monitoring data that can be used for fault detection, diagnosis, quantification, prognosis, and maintenance planning. One commonly used method involves predicting power generation using SCADA data and comparing it against measured power generation. Significant discrepancies between predicted and measured values can indicate sub-optimal operation, natural aging, or unnatural faults.

Various predictive models, including parametric and non-parametric (statistical) approaches, have been proposed for estimating power generation. However, the imperfect nature of these models introduces uncertainties in the predicted power output. Additionally, SCADA monitoring data is prone to uncertainties arising from various sources. The presence of uncertainties from these two sources – imperfect predictive models and imperfect SCADA data – introduces uncertainty in the predicted power generation. This uncertainty complicates the process of determining whether discrepancies between measured and predicted values are significant enough to warrant maintenance actions.

Depending on the nature of uncertainty – aleatory, arising from inherent randomness, or epistemic, stemming from incomplete knowledge or limited data – different analytical approaches, like Probabilistic and Possibilistic, can be applied for effective management. Both, Probabilistic and Possibilistic, Approaches offer distinct advantages and limitations. The Possibilistic Approach, rooted in fuzzy set theory, is particularly well-suited for addressing epistemic uncertainties, especially those caused by imprecision or sparse statistical information. This makes it especially relevant for applications such as wind turbines, where it is often challenging to construct accurate probability distribution functions for environmental parameters due to limited sensor data from hard-to-access locations.

This research focuses on developing a methodology for identifying sub-optimal operation in wind turbines by comparing Grid Produced Power (Measured Produced Power) with Predicted Produced Power. To achieve this, the paper introduces a Possibilistic Approach for power prediction that accounts for uncertainties stemming from both model imperfections and measurement errors in SCADA data. The methodology combines machine learning models, used to establish predictive relationships between environmental inputs and power output, with a Possibilistic Framework that represents uncertainty through possibility distribution functions based on fuzzy logic and interval analysis. A real-world case study using operational SCADA data demonstrates the approach, with XGBoost selected as the final predictive model due to its strong accuracy and computational efficiency.

Conflict of Interest Statement

The author declares no conflicts of interest.