Artificially intelligent models for the site-specific performance of wind turbines

Abstract

Power developed by the wind turbines, at different wind velocities, is a key information required for the successful design and efficient management of wind energy projects. Conventionally, for these applications, manufacturer's power curves are used in estimating the velocity-power characteristics of the turbines. However, performance of the turbines under actual field environments may significantly differ from the manufacturer's power curves, which are derived under 'standard' conditions. In case of existing wind projects with sufficient performance data, the velocity-power variations can better be defined using artificially intelligent models. In this paper, we compare the performance of four such models by applying them to a 2-MW onshore wind turbine. Models based on ANN, KNN, SVM and MARS were developed and tested using the SCADA data collected from the turbine. All the AI models performed significantly better than the manufacturer's power curve. Among the AI methods, SVM-based predictions showed the highest accuracy. A site-specific performance curve for the turbine, based on the SVM model, is presented. Wider adaptability of this approach has been demonstrated by successfully implementing the model for a 3.6-MW wind turbine, working under offshore environment. Being "site-specific data" driven, the proposed models are more accurate and hence better choice for applications like short-term wind power forecasting and pro-diagnostics of wind turbines.

Estimation of Wind Turbine Performance Degradation with Deep Neural Networks

“…Study by Kim and Kim (2021), based on the SCADA data, uses the turbine's power curve for performance comparison. As discussed in Veena, Mathew and Petra (2020), limitations of manufacturer's power curve in understanding the site specific dynamics of the velocitypower response of the turbines has been well established in several previous studies. Further, the analysis is based on four years performance of the turbines, which may not be sufficient to capture the time series performance degradation.…”

“…Wind turbines have two distinct operational regions viz. the dynamic region corresponding to the cut-in to rated wind velocities, and the deterministic region corresponding to the rated to cut-out velocities (Veena et al, 2020). Out of these, performance of the turbine between the cut-in and rated wind velocities were considered for this study.…”

“…In Veena et al (2020), four different algorithms have been used to estimate the performance of a similar wind turbine of 2 MW size in a different location. Artificial neural network (ANN), support vector machine (SVM), k-nearest neighbors (k-NN), and multivariate adaptive regression splines (MARS) algorithms were used for this purpose.…”

Power production forecast for distributed wind energy systems using support vector regression

“…Due to the stochastic nature of the wind, power output from the wind turbines can fluctuate significantly, even within short time intervals. 2 To efficiently manage such grids, variations in energy contributed by the turbines must be quantified in different time scales. Accurate wind power forecasts must understand these power fluctuations and manage the resulting uncertainties.…”

“…Wind power production forecasting can generally be classified as physical methods, traditional statistical methods, and, more recently, the so-called learning methods (e.g., machine learning [ML] approaches). [2][3][4][5][6] ML methods are considered an alternative to conventional methods as they have shown their ability to accurately predict wind power production. 6 One or more of these approaches can be combined to develop hybrid forecasting methods.…”