Wind Farm Control Nowcasting
What is nowcasting
The term Nowcasting means fore-casting for a relative short prediction horizon. For wind farm applications short is below some hours.
What is it good for
It is important for the electricity grid and thus the transmission system operator (TSO), to know what a farm will produce in the next hour [Giebel and Kariniotakis, 2009, Landberg, 2012]. The Danish TSO point to 10 min as the most relevant prediction horizon [Pinson, 2012]. The uncertainty on the forecast is as important as the point forecast itself [Landberg, 2012]. If no uncertainty is provided then for example the probability of lag of power in the next hour will be unknown. The inter hour forecast can also be used by the farm owner/operator for trading in the balancing market.
A medium range forecast for horizon’s from 12-36 hours can be used for energy trading e.g. at Nord Pool.
From meteorology we have models used for weather forecast. The models are based on the laws of physics and are very power full for forecast from one to approximately 10 days ahead [Shaojin et al., 1991]. Forecasts for more than 10 days e.g. one month should be based on seasonal statistics as the physical models loses their power on these long horizons [Lorenz, 1963]. For short horizons e.g. less than one hour the meteorology models becomes less useful. For very short horizons like one minute, statistical models using available measurement will be superior. Moreover, persistent methods [Glassley et al., 2010, Giebel and Kariniotakis, 2007] where the most recent measurement is used as the prediction is close to optimal. However a good method for the uncertainty are still needed. For horizons increasing from seconds to days the focus should move from statistical to meteorological methods and models [Glassley et al., 2010]. These two concepts are illustrated in figure 1.
The models developed here are not using any meteorological forecasts. They only use standard measurement which are available in commercial wind farms. The models are based on standard statistical time series analysis. The atmospheric conditions as 10 minutes average wind speed and direction, turbulence intensity and atmospheric stability is typically only varying slightly within the hour but often changes dramatically over hours and days. Consequently some form of adaptive methods must be used to follow these changes. For this standard forgetting factor methods for system identification is used for two simple models which predicts total wind farm power from its historical values.
Persistence based on WF power including uncertainty
If the signal in question, here wind farm power, is called the persistence method is by definition
From a time series point of view this is consistent with using a Wiener process for
where denotes a Wiener process in continuous time with incremental variance . By definition this means that is normal with independent increments with variance proportional to time lag. From this the predictor and its prediction error variances follows from (3) below which holds for any .
The only parameter in this model is . As it is not known it has to be estimated.
Beyond persistence models
In principle all time series model structures could be used instead of the persistence model. For example the simple first order discrete time (AR) process can be used.
The parameters , , can then be estimated in a adaptive/recursive way. Also a approximation for the confidence limits can be found. Strictly speaking, AR processes has no input and therefore has mean value 0. The term can be interpreted as i.e. a input which is always 1. In this sense the model type is an auto regressive model with exogenous input (ARX). The parameter only serves to give a non zero mean value (5) therefore the model is still called a AR model.
Alternative methods and models
It is very important to consider exactly what inputs are used was outputs are predicted and what assumptions are done. For example, is it total wind farm power that is predicted assuming the wind farm always runs at max power or is the farm allowed to be derated and then not the produce power but the available power is predicted.
There are a number of other choices that defines the forecasting method. The most important are:
- Start with wind speed and direction and map to power or directly use power.
- Use values from single turbines or aggregated values from wind farm.
- Model type used: persistence, AR etc.
Any of these many combinations could potentially be a good model. Methods using power of course only works when derating is not used. Some of the most reasonable models are presented and discussed in more details in Knudsen .
-  G. Giebel and G. Kariniotakis. Best practice in short-term forecasting - a users guide. In Conference proceedings (online). EWEA, European Wind Energy Association (EWEA), 2007.
-  G. Giebel and G. Kariniotakis. Best practice in short-term forecasting - a users guide. Technical report, Risø National Laboratory for Sustainable Energy, DTU, Denmark, 2009.
-  W. Glassley, J. Kleissl, C. C. van Dam, H. Shiu, J. Huang, G. Braun, and R. Holland. California renewable energy forecasting, resource data and mapping. Reports 500-99-013, California Institute for Energy and Environment, 2010. Appendix BWind Energy Forecasting: A Review of State-of-the-Art and Recommendations for Better Forecasts.
-  T. Knudsen. Wind farm power nowcasting for less than an hour - offwind project. Technical report, Automation and Control, Department of Electronic Systems, Aalborg University, 2014.
-  L. Landberg. Taking the guesswork out of wind power forecasting. In EWEA2012, Monday 16 - Thursday 19 April 2012, Copenhagen, Denmark. EWEA, EWEA, 2012. Slide presentation.
-  E. N. Lorenz. Deterministic nonperiodic flow. Journal of the Atmospheric Sciences, 20(2):130–141, 1963. doi: 10.1175/1520-0469(1963)020<0130:dnf>2.0.CO;2.
-  P. Pinson. Very-short-term probabilistic forecasting of wind power with generalized logit-normal distributions. Journal of the Royal Statistical Society, Appl. Statist., 61:555–576, 2012. doi: 10.1002/for.1194.
-  Y. Shaojin, P. Yongqing, and W. Jianzhong. Determination of kolmogorov entropy of chaotic attractor included in one-dimensional time series of meteorological data. Advances in Atmospheric Sciences, 8 (2):243–250, 1991. doi: 10.1007/BF02658098.