De kommende tre åra skal jeg jobbe med en PhD innen fornybar energi/ energimeteorologi/ vindkraft/ anvendt statistikk. De siste månedene har gått med til å utarbeide en mer detaljert prosjektbeskrivelse, gjengitt i lett forkorta utgave under. Å planlegge hva en skal holde på med i tre år, spesielt innenfor et emne en strengt tatt foreløpig ikke kan veldig mye om, har vært krevende, og det blir spennende å se i hvor stor grad planen faktisk vil bli fulgt. Uansett, jeg synes det bare virker mer og mer spennende, og gleder meg til fortsettelsen.
Background
Recent years there has been an increased focus on wind as a source of energy, which has resulted in a significant increase in installed production capacity and market share. With today’s strong political focus on climate change and becoming less dependent on fossil fuels this tendency is unlikely to change. According to the European Wind Energy Association (EWEA) the total installed wind power in the EU raised from ~13 GW in 2000 to ~65 GW in 2008 (22,3 % average annual increase), with a target of further increase to 230 GW in 2020, corresponding to a market share of 20 %. The development in Norway have had a slightly slower pace, with an increase in the installed wind power from 97 MW in 2002 to 525 MW in 2011 (20,6 % average annual increase).
However, the increased share of wind power also raises some challenges. Unlike conventional power plants, the production of wind power is to a large extent dependent on factors beyond human control, most important the magnitude of the wind. Accordingly this makes the level of production less stable and harder to adjust to the actual needs. A secure electricity supply requires that the electricity production mirrors the consumption as exactly as possible. The power consumption, and its variations over the day, is from experience rather well known. It is therefore possible to estimate the power consumption for the next day with a high accuracy and to adjust the production according to this. Large shares of wind energy make this more difficult, as it is no longer just the consumption that is subject to variability. On top of the variability of the wind comes the fact that the conversion from wind speed to power output is highly non-linear. Hence, the power curve inflates or deflates the prediction error of the wind speed according to its local derivative.
Another major difference between wind and most conventional sources of energy is that wind cannot be stored. The consequence of this is that if wind power production is stabilized through not utilizing the full production-potential, energy equivalent to the reduction is lost. This again reduces the profitability of the wind farm.
Wind power forecasting is a collective term for methods to predict future wind farm power output. Short-term WPF is recognized as an important requirement by most involved in the wind power industry. The accuracy of a prediction model is of primary concern as it is directly linked to the security of supply and to the operation cost. The use and potential benefits of WPF depend on the time scale considered. Very short-term forecasts, forecasts of up to a few minutes, can amongst other be used for the turbine active control and thus both reduce the structural stress and optimize the power production. Forecasts for up to 48-72 hours are amongst other useful for the Transmission System Operator (TSO) and for energy trading. Reliable forecasts in this time-span will enable TSOs to optimize the use of control power, and thereby be both more environmental friendly and more cost efficient. Holttinen and Parkes et al. has also shown that WPF can contribute to increasing the economic value of wind energy. WPF for longer timescales than 72 hours may be of interest for planning of maintenance of the wind farms, transmission lines etc.
Little work has been done on wind power predictions for the Norwegian energy market. In the most substantial one statistical and one physical model are tested on data from Vikna wind farm. In this it is stated that it is their general impression that the Norwegian wind energy market is in its infancy with regards to the use of wind power predictions. The reasons for this is believed to be that the production is too small for the energy companies to invest in the forecasts, thus it is likely to believe that the demand for WPF will increase as the wind power production increases.
Overall aim
· To collect and pre-process data form Norwegian wind farms making it suitable for running and evaluating different methods for WPF.
· To gain increased knowledge of how spatial smoothing effects influence the WPF error for ensembles of wind farms.
· To gain increased knowledge of the spatial-temporal propagation of WPF error for wind farms.
· To develop a method to estimate the uncertainty of WPF for ensembles of wind farms.
· To develop a method for WPF suited for “Norwegian” conditions.
Basic considerations
The present project is concerned with improving the precision of WPF for individual wind farms and for ensembles of wind farms, and with quantifying the uncertainty of the forecasts. In the previous sections it has been shown that reliable WPF are a requirement for maintaining a secure electricity supply with a high penetration of wind power. It has also been shown that WPF in itself might contribute to increasing the economic value of wind power.
As shown above, the majority of the models for WPF in the literature are developed with data from Danish or North-German wind farms. This must be regarded a natural consequence of the fact that these are the areas with the highest penetration of wind power, thus these are the areas where the incentives for creating reliable models are strongest. Nevertheless, both Denmark and North-Germany substantially differ from the coastal regions in Norway when it comes to topology and terrain-complexity. It is therefore not given that these models will be equally well-performing when applied on Norwegian wind farms.
Kariniotakis et al. investigated the relation between model performance and site characteristics. It was found that the performance of the power prediction models investigated decreased with increasing complexity of the terrain. It was also found that the variation between the performances of the models increased with increasing complexity, thus making choosing the model best suited for the local conditions important.
An interesting aspect of Norwegian wind farms is the enormous north-south spread. It is approximately the same distance between the southernmost and the northernmost wind farms in Norway as it is from the southernmost wind farm to Rome in Italy. A consequence of this is that even the large-scale weather situation might differ, and that the covariance structures between the different wind farms might be considerably weaker than in the examples presented in the literature.
Wind power production data and on-site wind measurements from wind farms are considered business sensitive, and are therefore hard to obtain. Presently I have data from two wind farms available and I am in a promising dialogue with the owners of two more wind farms. In addition to this, to increase the amount of data for model building and testing, I plan to use data from Met.no meteorological measuring stations. These are easily accessible on-line, and in general of very high quality. The main difference between these data and the data from actual wind farms is that they don’t reflect the effect of the power curve. However, for evaluation and comparison of different models, the effect of the power curve can be obtained through transformations.
Planned research directions and research topics
I argue that the complexity of the Norwegian terrain entail that the most commonly used methods of WPF from the literature will not have an optimum performance when applied to Norwegian wind farms. Moreover I argue that the potential error smoothing effects from ensembles of wind farms will be larger for Norwegian wind farms than in most examples given in the literature. Through analysis of empirical data I will single out the challenges and shortcomings of existing methods. This knowledge will be used to making improved methods more suited for Norwegian conditions.
Based on the research directions described above, the following topics have been identified and will be addressed at the different phases of this PhD-project.
- Empirical post-processing of data
A natural starting point of the thesis is to gain knowledge of the actual situation. As earlier mentioned, little work has been done on WPF or analysis of WPF errors from Norwegian wind farms, and a clear description of the magnitude of the problem has therefore not been established. It is however well documented that the annual wind power production the last years consistently has been lower than expected. Parts of this can be explained by lower wind farm availability than expected, but there is also an unexplained negative deviation.
The empirical post-processing of the data aims at identifying factors (wind direction, wind speed etc.) that cause deviance between predicted and observed wind speed, and to quantify the impact of these factors. This will amongst other be done by partitioning of the data, calculation of describing parameters and, where data is available, calculation of conditional empirical power curves.
From the empirical post-processing of data I hope to gain insight in which sources are the main contributors to variation in power production. I also hope to be able to explain and quantify the amount of variation caused by these factors.
- Smoothing effects for ensembles of wind farms
Spatial smoothing effects is a well-established fact, and it has been shown that the magnitude of the error reduction from smoothing effects mainly depends on the size of the region under consideration. Different models have been proposed to generalize the smoothing effect for different ensembles of wind farms. Most models however, have only been tested for wind farms in less complex terrain and relatively small and homogenous regions.
Closely linked to the smoothing effects is the spatial-temporal propagation of forecast errors. Studies of the spatial-temporal propagation of forecast errors involve investigating the correlation structures between the forecast errors of different wind farms with various time-lags. These can amongst other be used as a correction to the NWP forecasts for wind farms downwind of the wind farm from which the forecast errors are measured.
The work on smoothing effects will consist of three rather distinct parts. First I will utilize well-established methods from the literature to calculate the error reduction obtained from smoothing effects. The results from this will be compared to the effects obtained by the same models from other areas. Then I will attempt to improve the results by making adaptions to the best performing models. Finally I will try to generalize the results, and propose a general model for estimating the error reduction obtained from smoothing effects.
The work on smoothing effects aims at quantifying the obtainable error reduction for various regions and ensembles of wind farms, and to uncover the underlying structures that causes the error reductions. It is further a goal to propose a general method for choosing sample wind farms.
- Uncertainty of estimates for ensembles of wind farms
WPF is usually given as point-forecasts, and provide the single value that is the most likely outcome. Nevertheless, it is a fact that the probability that exactly the value from the forecast occurs is close to zero. A point forecast is always subject to an error. When unknown, the level of error inhibits an optimum use of WPF. To utilize the full potential of WPF it is therefore necessary to also have an estimate of the uncertainties involved. Often this is given as an interval-estimate, for instance limited by the upper and lower quantile.
Numerous methods for estimating the uncertainty of WPF for single wind farms have been proposed, spanning from parametric methods where the prediction errors are fitted to a known distribution, to methods based on ensemble NWP. Parametric methods derive the uncertainty from past experience, while ensemble NWP gives an image of the uncertainty involved in the wind prediction. However, little work is yet done on estimating the uncertainty of WPF for ensembles of wind farms.
The uncertainty of WPF for ensembles of wind farms obviously will depend on the level of uncertainty of the individual wind farms. In addition it will be influenced by the covariance structure of the wind farms in the ensembles and by the spatial-temporal propagation of forecast errors. A natural starting point for the work on uncertainty of estimates for ensembles of wind farms will be to try to extend methods developed for single wind farms. I will also investigate the possibility of giving a theoretical description of the sum of the individual distributions, and of describing the sum of distributions by simulation-centred approaches.
The work on uncertainty of estimates for ensembles of wind farms aim at describing the distribution of the regional power production estimate, thus the distribution of the sum of the power production estimates from the individual wind farms. It is an objective to generalize the method of estimating uncertainty, and to integrate this with the method proposed for choosing sample wind farms.
