1.2 Previous research
Developing models for short-term prediction of wind power production is by no means a trivial task, as the underlying system covers everything from the large scale atmospheric flow, influence by local topography, veg-etation and atmospheric conditions, the wind farm layout and the single turbine. This system, including each single component, is by nature non-linear and non-stationary.
Short-term prediction of wind farm power production has already been the subject of extensive research prior to this study. The approach used in this research can be distinguished by the type of input-data used in the prediction models. In principle there are three categories; models based on local measurements, models based on numerical weather predictions, and finally, models based on a combination of both local measurements and numerical weather predictions. The following sections provide an overview of these types of models and bibliographic notes.
As mentioned previously, the research presented in this thesis has been performed in a collaboration between the Department of Informatics and Mathematical Modelling at the Technical University of Denmark and the Research Programme Wind Power Meteorology at Risø National Labora-tory. These departments have been working within the short-term wind power prediction field for a long time, Risø since 1989 and the Techni-cal University of Denmark since 1992. Both departments have partici-pated, and still do, in international research projects related to this field.
This research has resulted in two on-line software prediction systems, im-plementing short-term prediction models which today are considered as state-of-the-art.
1.2.1 Local measurements
The methodologies that have been applied to local measurements are within the field of time series analysis, regression analysis and neural networks. One of the easiest prediction models is the persistence model.
In this model, the prediction for all prediction horizons is set to the most recent measurement value. This means, by definition, that the error for
the now cast, i.e. zero prediction horizon, is zero. Furthermore, for short prediction horizons, i.e. on the order of minutes or a few hours, the error is relatively small compared to the errors for predictions from numerical weather prediction models or more sophisticated time series models. This is because the atmosphere is quasi-stationary, the time scales in the atmosphere are in the order of days (at least in Europe).
It takes about one to three days for a low-pressure system to cross the continent, high-pressure systems can be more stationary. As the pressure systems are the driving force for the wind, the changes in the wind have time scales of the same order. Therefore, the persistence model has been used as a comparative model for other prediction models. For longer prediction horizons, i.e. more than a few hours, the persistence model is not adequate as a comparative model. In Paper Cit is shown that a first order auto-regressive model is more adequate.
In the research described by (Bossanyi 1985) a Kalman Filter with the last 6 values (1 minute averaged data) as input is used to predict the next step. This gave 10 % improvement in the RMS error compared to the error of the persistent predictions for the next time step. This improvement decreased for longer averages, and disappeared completely for 1-hourly averages.
In (Dutton, Kariniotakis, Halliday & Nogaret 1999) an autoregressive model and an adaptive fuzzy logic based model for the cases of Crete and Shetland showed minor improvements over persistence for 2-hour horizons. For longer horizons significant improvements were found, i.e.
for the 8 hour horizon an 20 % improvement in the RMS error was found.
However, as described in PaperC, the persistence model is not adequate as reference for these horizons. Furthermore, the fact that most of the likely wind speeds were contained in the 95 % confidence band for the longer horizons, means that using the mean value for all times as the predictand would provide almost the same RMS error results compared to the models in this paper.
The early models (Madsen 1996) developed at the Department of Infor-matics and Mathematical Modelling, were based on local measurements only. Like the models described above, these models did not perform well for prediction horizons above 6–12 hours.
1.2 Previous research 7 In (Beyer, Degner, Hausmann, Hoffmann & Ruj´an 1994) neural networks are used for next-step prediction of either 1-minute or 10-minute aver-aged data. In both cases they find 10 % improvement over the persistence model. This is achieved with a rather simple topology, while more com-plex neural network structures did not improve the results further. In (Tande & Landberg 1993) it is found that neural networks used to predict 10-second values using 1-second averages perform only marginally better than the persistence model. In (Bechrakis & Sparis 1998) neural net-works are used to utilise wind direction information, but no performance measures over persistence are presented.
1.2.2 Measurements and Numerical Weather Predictions Based on the methodology developed for the European Wind Atlas (Troen
& Petersen 1989), Risø National Laboratory has developed short-term prediction models based on physical reasoning (Landberg 1999, Landberg
& Watson 1994). These models are primarily based on numerical weather predictions as input, and are more thoroughly described in Chapter 2.
The University of Oldenburg (Beyer, Heinemann, Mellinghoff, M¨onnich
& Waldl 1999) has developed models similar to those developed at Risø.
The main difference is that the models developed at Oldenburg use nu-merical weather predictions from the Deutschlandmodell of the German Weather Service DWD instead of HIRLAM.
Vitec AB from Sweden is working on a model based on meteorological forecasts from Swedish Meteorological and Hydrological Institute SMHI.
So far, nothing is published (Giebel 2000).
In (Martin, Zubiaur, Moreno, Rodriguez, Cabre, Casanova, Hormigo &
Alonso 1993) a prediction tool for the rather special case of Tarifa/Spain is described. Due to the special topological situation for the wind farms in Strait of Gibraltar, they could predict the power output from the pressure difference between measurements at Jerez and Malaga airports, with the additional use of Spanish HIRLAM. The founding for this project was stopped, and the project therefore ended half way through.
As mentioned previously, the early models developed at the Department
of Informatics and Mathematical Modelling, were only taking local mea-surements as input. Therefore, models which included meteorological forecast were developed (Nielsen & Madsen 1996), and it is a statistical model which takes both measurements and meteorological forecasts as input, which is implemented in the on-line prediction system, WPPT (Wind Power Prediction Tool), developed at this department. This sys-tem is briefly outlined in PaperH.
The models developed at Risø National Laboratory, also make use of local measurements. This is mainly for calibration purposes, also described as MOS (Model Output Statistics) (Landberg & Joensen 1998). Therefore it is important to note that on-line measurements are not used in this model to calculate the actual predictions.
EWind is an US-American model by TrueWind, Inc (Bailey, Brower
& Zack 1999). Instead of using a once-and-forall parameterization for the local effects, like the Risø approach does with WAsP, they run the ForeWind numerical weather model as a meso-scale model using bound-ary conditions from a regional weather model. Due to the enhanced resolution in the meso-scale model more physical processes are captured, and the predictions can be better tailored to the local site. Nevertheless, they use adaptive statistics to remove the final systematic errors. No performance results are presented.
In (Shuhui, Wunsch, O’Hair & Giesselmann 2001) regression and neu-ral network methodology is compared in the aim of modelling the wind turbine power curve. From the models tested it is concluded that the neural network approach is superior to regression. The power curve is es-timated using local measurements of meteorological variables and power.
This power curve is then supposed to be used for the transformation of numerical weather predictions to predictions of the power production.
This approach is not sound as the properties of the numerical weather prediction are not necessarily the same as the properties of the measure-ments, i.e. properties like the statistical metrics mean and variance. This problem is further described in (Jonsson 1994), which argues that if er-rors are present in the regressors the use of the true system for prediction will not result in optimal predictions. No results from using numerical weather predictions are presented in this paper.