Short Term Wind Turbine Power Output Prediction

Abstract

In the wind energy industry, it is of great importance to develop models that accurately forecast the power output of a wind turbine, as such predictions are used for wind farm location assessment or power pricing and bidding, monitoring, and preventive maintenance. As a first step, and following the guidelines of the existing literature, we use the supervisory control and data acquisition (SCADA) data to model the wind turbine power curve (WTPC). We explore various parametric and non-parametric approaches for the modeling of the WTPC, such as parametric logistic functions, and non-parametric piecewise linear, polynomial, or cubic spline interpolation functions. We demonstrate that all aforementioned classes of models are rich enough (with respect to their relative complexity) to accurately model the WTPC, as their mean squared error (MSE) is close to the MSE lower bound calculated from the historical data. However, all aforementioned models, when it comes to forecasting, seem to have an intrinsic limitation, due to their inability to capture the inherent auto-correlation of the data. To avoid this conundrum, we show that adding a properly scaled ARMA modeling layer increases short-term prediction performance, while keeping the long-term prediction capability of the model. We further enhance the accuracy of our proposed model, by incorporating additional environmental factors that affect the power output, such as the ambient temperature and the wind direction.

The authors acknowledge the Daisy4offshore consortium for the provision of the data. The work of Sándor Kolumbán and Stella Kapodistria is supported by NWO through the Gravitation-grant NETWORKS-024.002.003. Sándor Kolumbán also received funding from the Institute for Complex Molecular Systems (ICMS) in Eindhoven. Furthermore, the research of Stella Kapodistria was partly done in the framework of the TKI-WoZ: Daisy4offshore project.

A Appendix: Data

The goal of this section is to describe the features of the data used in this study. The data was obtained by the supervisory control and data acquisition (SCADA) system of a wind turbine operator in the Netherlands. The data was collected from an off-shore Vestas V80-2.0MW wind turbine, with a rated capacity of 2MW. Vestas V80-2.0MW joins the grid connection at a wind speed of 4 m/s, has a rated actual power output of 2 MW (typically achieved) at a wind speed of 16 ms, and it is disconnected at a wind speed of 25 ms. See Fig. 1 for a depiction of the theoretical WTPC. These are suggested values offered by the manufacturer, but they might change due to wear of the turbine or due to installation or geographic circumstances.

The data used for the analysis presented in this paper spans across two years and the dataset contains recordings of the environmental conditions, as well as the physical state, and power output of the turbine.

There are two important features of SCADA datasets, which are not specific to the data of our study but are common amongst SCADA datasets recorded throughout the wind industry. One of these is the 10 min reported frequency of the SCADA observations; although the signals of interest are collected at a relatively high frequency, only processed observations calculated on a 10 min window are recorded in the SCADA databases. These processed signals contain the average, maximum, minimum and standard deviation of the wind speed, and the power output amongst other quantities of interest. The second important feature is that the data are strongly quantized due to the rounding of the reported number. As a result, the observations are recorded up to one decimal digit. Some of our finding are consequences of these two properties which correspond to the quasi industry standard. Because of this, we expect that our results are also applicable to similar data coming from other wind turbine operators or wind turbine service providers.

1.1 A.1 Description of the Raw Dataset

All graphs and figures were produced using two seasonal parts of the available dataset. Throughout the paper, we refer to the data recorded between June 1, 2013, and August 31, 2013, as the training data, and the corresponding period of year 2014 as the validation data. Although, we have access to the full two year data set, we choose to restrict our analysis in a specific season of the year, as this reduces seasonality effects, while still maintaining a significant amount of data, and it permits a full decoupling between the training and the validation data. It is important to note that the results presented in the paper still hold when we perform the same analysis using the full year 2013 as training data and the data from 2014 for validation.

The dataset contains observations of various signals every 10 min. Some of the signals contained in the dataset are the ambient wind speed, say $w_t$, the relative direction of the wind speed with respect to the nacelle, say $\phi _t$, the ambient temperature, say $T_t$, and the power output produced by the turbine, say $p_t$, at time t, $t\ge 0$. Besides the aforementioned continuous valued signals, there are some nominal variables with a discrete support, such as the variable pertaining to the different operational states of the turbine. Such variables help to identify time periods during which the turbine is out of use (maintenance, free run, blades turned into low resistance position) or if the wind turbine is in a state different from normal operational condition.

In the first part of the paper, we suppress the subscript t as we deal with static models, while in the second part of the paper we deal with dynamic models, and we, therefore, reinstate the subscript t notation.

Table 8. Summary report of the data cleaning

Fig. 18.

The power output, p, against the wind speed, w, in the raw (red) and the cleaned (blue) dataset from 2013 (Color figure online)

1.2 A.2 Cleaning the data

The quality of the available SCADA data is extremely good, nevertheless it requires some pre-processing before creating the forecasting models. We list below the cleaning rules implemented in this study, according to which we disregard observations:

• 1. Missing entries (NAs): there are a few timestamps that are completely missing from the 10 min sampling sequence.

• 2. Incomplete entries (IN): if one or more of the signal values, e.g., the power output, the wind speed, etc., are missing from a data record, then the full record corresponding to this time stamp is discarded.

• 3. Not normal operation (NNO): based on the value of the state variables we can disregard states that do not correspond to normal operational conditions, e.g. free rotation of the wind turbine without connection to the grid, derated operation, etc.

• 4. Outliers: Firstly, all observations of wind power corresponding to the same wind speed are grouped together and the corresponding box plot is generated. Then, for every given wind speed value, all points with power generation outside the whiskers of the box plot (i.e., all observations falling outside the interval $(Q_1-3\text {IQR},Q_3+3\text {IQR})$) are discarded.

Table 8 contains the summary report of the data cleaning procedure. It shows that approximately 5% of the original data is discarded, still leaving a trove of data to be used for estimation purposes. The scatter plot of the power output, p, against the wind speed, w, is shown in Fig. 18. In this figure, we have color-characterized the training data by depicting in red the raw data, and in blue the cleaned dataset used for the analysis.