South East Australia Wind Power Study - CiteSeerX
SOUTH EAST AUSTRALIA WIND POWER STUDY Robert Davy and Peter Coppin Wind Energy Research Unit CSIRO Atmospheric Research
100 90 80 Capacity %
70 60 50 40 30 20 10 0 11-Jun-01
13-Jun-01
15-Jun-01
17-Jun-01
19-Jun-01
21-Jun-01
This report was prepared for the Australian Greenhouse Office by the CSIRO Wind Energy Research Unit, October 2003
Published by the Australian Greenhouse Office. © Commonwealth of Australia 2003 ISBN: 1 920840 96 6
This work is copyright. It may be reproduced in whole or in part for study or training purposes subject to the inclusion of an acknowledgment of the source, but not for commercial usage or sale. Reproduction for purposes other than those listed above requires the written permission of the Australian Greenhouse Office. Requests and inquiries concerning reproduction and rights should be addressed to: The Manager Communications Australian Greenhouse Office GPO Box 621 CANBERRA ACT 2601
IMPORTANT NOTICE – PLEASE READ This report has been prepared by the CSIRO for the Australian Greenhouse Office. This report is produced for general information and does not represent a statement of the policy of the Australian Government. While every effort has been made to ensure accuracy and completeness, no guarantee is given, nor responsibility taken by the Australian Government for errors or omissions in the report, and the Australian Government does not accept responsibility in respect of any information or advice given in relation to or as a consequence of anything contained here. The Australian Government and all persons acting for the Australian Government in preparing this report accept no liability for the accuracy of or inferences from the material contained in this publication, or for any action as a result of any person’s or group’s interpretations, deductions, conclusions or actions in relying on this material.
Important CSIRO Disclaimer To the extent permitted by law, CSIRO Atmospheric Research (including its employees and consultants) excludes all liability to any person for any consequences, including but not limited to all losses, damages, costs, expenses and any other compensation, arising directly or indirectly from using this publication (in part or in whole) and any information or material contained in it.
Contents Summary ............................................................................................................................. 1 Study Purpose and Scope .................................................................................................... 3 Introduction ......................................................................................................................... 3 Data Preparation .................................................................................................................. 4 Analysis............................................................................................................................... 7 Analysis of Periodic Components of Wind Power............................................................ 17 Examples of Wind Events in South East Australia ........................................................... 19 Wind and NEM Load ........................................................................................................ 22 References ......................................................................................................................... 24
Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9
Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19
Location of Automatic Weather Stations used in the study- colours represent topographic height in metres ........................................................................... 4 Power curve for Vestas V80-2.0MW,104.0db(A) at sea level....................... 6 Generation duration curves by state ............................................................... 8 Generation duration curves by season............................................................ 9 Generation duration curves by state for hours 8am to 8pm inclusive .......... 10 Generation duration curves by state for hours 9pm to 7am inclusive .......... 10 Reliably available capacity figures by state and times of day...................... 11 Frequency of extreme power production events (extreme low: output95% of installed capacity)............ 12 Frequency of observations at Goulburn AWS (bars) and a Weibull fit (line) showing a distribution more typical of wind energy sites. Gaps in the distribution are due to the original measurements being taken in integer knots. ............................................................................................................. 12 Effect of wind farm aggregation on average length of low power events ... 13 Effect of wind farm aggregation on average length of high power events ... 13 Average power by month of year and by state.............................................. 14 Hourly incremental change density by state.................................................. 15 Cumulative variance of power output as wind farm areas are combined. Time period 24 Jun 2002 to 18 Sep 2002 (“Winter”)............................................. 18 Cumulative variance of power output as wind farm areas are combined. Time period 05 Dec 2001 to 18 Mar 2001 (“Summer”)......................................... 18 Estimated power production - cold fronts during winter .............................. 19 Estimated power production - diurnal cycle in summer................................ 20 Estimated power production – extended calm in South Australia (winter) .. 20 Estimated power production – strong winds in NSW (summer)................... 21
Summary Whilst wind farm development has been accelerated in Australia by the Mandatory Renewable Energy Target (MRET), the majority of proposed wind farms on the mainland to date are located in a restricted number of areas, with particular interest evident in South Australia. This report has been undertaken to investigate the possible benefits of distributing wind farms across the south east of the continent, including both coastal and inland sites in South Australia, Victoria and New South Wales. Using modified Bureau of Meteorology wind data from 9 sites over 4 years, the results of this analysis indicate that more widespread distribution of wind farms (or greater locational ‘diversity’) would result in the following benefits: •
reduction in frequency and average length of extreme power events
•
reduced variation in the average contribution of wind energy
•
reduced volatility of average aggregated output
•
improved ability to forecast total wind power output
•
reduction in variance of total power on time scales up to about ten days long
However it was found that seasonal variations were similar across the south east of the continent and hence large scale aggregation would have little effect on the annual seasonal cycle in wind power output. These results indicate that variability in total wind power output can be reduced to some degree by wider distribution of numerous wind farms but remains substantial because synoptic patterns have a strong influence over the whole region. (Synoptic patterns are broad-scale wind patterns associated with the movement of low and high pressure systems across the continent.) Therefore, as in other countries forecasting would be a valuable tool in managing the impact of the variation. With wind farms and forecasting systems in place it will be possible to tune forecasts to measured wind farm output to improve forecast reliability and to develop appropriate confidence intervals according to the atmospheric conditions. Some correlation was found between the wind and load in the National Electricity Market (NEM) on a state-wise basis. This effect was observed for South Australia and Victoria but only for summer weather patterns. Further analysis would be required to draw useful conclusions on the match between regional load especially peak load and regional wind regimes. A large proportion of variation in load does not appear to be correlated with the variation in wind, and this suggests that forecasting of wind (together with standard load forecasting) will be useful for understanding wind power’s potential to efficiently displace non-renewable generation.
1
In a number of cases, results for New South Wales were limited by the use of Bureau of Meteorology data which were not wholly representative of wind energy sites on the Great Dividing Range. Note: It is important to remember that the results of this study are based on assumed state penetration levels and locational distributions based on current proposals. There has been no technically based analysis to date as to the levels of wind penetration that can be supported by any state. These factors mean that actual state penetration levels may ultimately be quite different from the levels assumed for this analysis. The results of this analysis depend on the assumed distributions and penetration levels and cannot necessarily be extrapolated if the actual distributions and capacities of wind farms in the south east of Australia differ.
2
Study Purpose and Scope The purpose of the work undertaken in this report is to better understand correlations and relationships between wind regimes across a number of states and investigate the effects of aggregating wind power across a wide area in south east Australia. The aspects covered include: •
the statistical availability levels of wind power based on the proportion of installed capacity that is statistically available for at least 95% of the time (referred to here as “reliably available capacity”);
•
the frequency of extreme wind events (lulls and high winds);
•
the duration of extreme wind events;
•
the spatial and temporal properties of the wind; and
•
the relationship with load in the National Electricity Market
Introduction Although wind farm developments to date have tended to be concentrated in a small number of areas, particularly coastal South Australia and Victoria, there are wind farm feasibility studies being undertaken throughout south eastern Australia where good wind resources have been identified. Aggregating wind farm output from diverse locations across the continent provides some ability to smooth out some of the variability in the supply of wind energy from any individual wind farm or single area. The weather systems which are the source of most of this broader scale variability generally move from west to east, taking times of the order of a day or more to traverse the full range of potential generation locations from western South Australia to northern NSW. The systems also evolve as they move and this together with a wide range of local effects allows wind generation in one location to be quite different to that in another. To investigate the extent of this spatial smoothing effect, a number of potential wind generation areas have been selected, the local wind energy generation calculated and statistics of the combined output examined.
3
Data Preparation Wind data were obtained from a number of Bureau of Meteorology Automatic Weather Stations (AWS) sites. The sites were chosen both to obtain a wide distribution across the three States and to broadly represent the areas of greatest current interest to wind farm developers in the south east of the continent, however AWS sites are not generally available for well exposed inland areas. The sites chosen for this analysis are shown in Figure 1 and were: •
South Australia: Coles Point, Sellicks Hill, Cape Jaffa
•
Victoria: Cape Nelson, Kilmore Gap, East Sale
•
New South Wales: Goulburn, Bathurst, Glen Innes
-28
Glen Innes
-30
-32
Bathurst -34
Coles Pt
Goulburn
Sellicks Hill -36
Cape Jaffa -38
Cape Nelson
KilmoreGap East Sale
1300 1200 1100 1000 900 800 700 600 500 400 300 200 100 0
-40
-42 134
136
138
140
142
144
146
148
150
152
154
Figure 1 Location of Automatic Weather Stations used in the study - colours represent topographic height in metres
4
Considering the simultaneous availability of full data sets from all AWS sites chosen, the need to use a whole number of years, and other data limitations, for instance Victorian sites were generally only available since 1999, the time span of the analysis was chosen to be 01 Mar 1999 to 01 Mar 2003, a four-year time interval. Some sites have not recorded half hourly data consistently over that time, swapping to hourly observations. Therefore, it was decided to use hourly observations throughout. Where samples were not taken on the hour the closest available sample was used instead (sometimes the nearest half hourly sample). Times were converted to Eastern Standard Time. Small gaps (up to 4 hours) were filled using linear interpolation. Some larger gaps were filled by combining two neighbouring stations - for example, some poor data at Bathurst spanning about one year were filled using Orange AWS. Also the Coles Point record was rather patchy and was filled using Port Lincoln. Several gaps in the data remained but it was preferred to leave them without further manipulation. This was so as not to artificially degrade or enhance any correlations which exist in the wind data. Overall the simultaneous availability of observations for the cleaned data was about 94%. Gaps were spaced in a fairly random fashion throughout the dataset, however the data availability did vary with calendar month. The lowest was May with 89% and the highest was April with 99%. This could have introduced a small amount of seasonal bias to those results which were not stratified by season, but generally these figures indicate that the data availability was sufficient for this kind of analysis. Wind speeds were scaled to yield overall averages of 8.0 ms-1 for inland sites and 8.5ms-1 for coastal sites as shown in Table 1. This scaling process was applied because 8.0 ms-1 is an approximate figure often regarded as the minimum for economic viability under the current regulatory circumstances, and coastal wind speeds are generally higher than inland (Coppin et al. 2003). Power estimates were obtained by applying the Vestas V802.0MW,104.0db(A) power curve as this represents a typical turbine being chosen for current developments. The turbine curve was interpolated using cubic splines. Appropriate air density figures were used for each site, as shown in Table 1, given its height above sea level. The turbine curve for sea level is shown in Figure 2. The turbine is designed to operate for wind speeds between 4 and 25 ms-1. The real response of a wind farm departs from this curve, firstly because it represents an average for a single turbine, and secondly because wind speeds are not the same across the whole wind farm site. The measured response of a wind farm would resemble the power curve in Figure 2, but would have some scatter and the cutoff at high wind speeds would be smoother. The use of a power curve therefore represents an approximation, and the simulated effect of the cutoff at high wind speed will be more pronounced than in any real system.
5
Table 1 Assumed air density and average wind speed for each site
Site Coles Point Sellicks Hill Cape Jaffa Cape Nelson Kilmore Gap East Sale Goulburn Bathurst Glen Innes
Scaled Average Wind Speed (ms-1) 8.5 8.5 8.5 8.5 8.0 8.5 8.0 8.0 8.0
Air Density (kg/m3) 1.21 1.18 1.21 1.21 1.15 1.21 1.12 1.09 1.09
2500
Power (kW)
2000 1500 1000 500 0 0
5
10
15
20
25
30
Wind Speed (m/sec)
Figure 2 Power curve for Vestas V80-2.0MW,104.0db(A) at sea level
Power output for each site was scaled to reflect an assumed installed capacity for the region as shown in Table 2. This step was intended to give a higher relative weighting to production in Victoria and NSW than existing proposals indicate at this stage in order to explore the impact of increased diversity in wind farm locations across the south east. However South Australian sites were still given the greatest weighting in relative terms reflecting the high number of proposed short-term wind developments and the excellent coastal wind resource. For this analysis, the relative weightings are more important than the absolute capacity levels.
6
Table 2 Weighting of wind farm capacity
SA (1600MW) Vic (1200MW) NSW (800MW)
Site Coles Pt Sellicks Hill Cape Jaffa Cape Nelson Kilmore Gap East Sale Goulburn Bathurst Glen Innes
Capacity MW 600 200 800 600 400 200 400 200 200
7
Analysis
Capacity %
Statistical availability The estimated power output from all regions was combined to analyse the way in which diversification of wind farm areas contributes to the proportion of total installed capacity which is statistically available more than 95% of the time. Cumulative histograms of the power output were calculated and plotted to produce generation duration curves shown in Figure 3 and Figure 4. The curves slope downwards from left to right, since 100% capacity is almost never available, and 0% capacity is almost always exceeded. In the generation duration plots the dotted black line is kept as a reference, being the curve for all sites at all times. 100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA
Vic
NSW
All
Figure 3 Generation duration curves by state
8
Capacity %
100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded Spring
Summer
Autumn
Winter
All
Figure 4 Generation duration curves by season
The results indicate that: •
when the wind power from all three states is combined, there is a tendency towards a flatter generation duration curve with a higher reliably available capacity (Figure 3).
•
the reliably available capacity is generally highest during summer and lowest during autumn (Figure 4).
•
higher reliably available capacity is observed during the daytime hours 8am to 8pm than during the remaining hours of 9pm to 7am which include overnight wind phenomena (Figure 5 and Figure 6).
9
Capacity %
100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA 8am to 8pm
Vic 8am to 8pm
All 8am to 8pm
All
NSW 8am to 8pm
Capacity %
Figure 5 Generation duration curves by state for hours 8am to 8pm inclusive 100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA 9pm to 7am
Vic 9pm to 7am
All 9pm to 7am
All
NSW 9pm to 7am
Figure 6 Generation duration curves by state for hours 9pm to 7am inclusive
Figure 7 shows that when wind power from the three states is combined the reliably available capacity increases considerably. However, the reliably available capacity across all states is less than the sum of the reliably available capacities from each state. It is interesting to compare this statistical quantity with “firm capacity”, a term commonly used to describe the availability levels for traditional fossil fuel generation that is able to be scheduled in the NEM. Wind energy has quite different uncertainty characteristics to
10
fossil fuel generation, which means that wind generation is not well described by the term “firm capacity”. Wind farm output is variable and not always dispatchable, as generation levels at a given time are determined by the variation in the wind resource. However, wind farms are potentially very technically reliable (i.e. little affected by equipment failure) because they contain a large number of relatively small generators. Thus for wind energy, predictability of the wind regime becomes the key network management issue. For these reasons this paper avoids referring to “firm capacity” as its meaning is often misinterpreted when used in relation to the statistical availability levels of wind energy. 14
% Capacity
12 10 8 6 4 2 0 SA
Vic 8am to
NS 9pm to
All All
Figure 7 Reliably available capacity figures by state and times of day
Frequency of extreme power events The frequency of extreme power production events (where an extreme low is defined as output of less than 5% of installed capacity and an extreme high is defined as output of greater than 95% of installed capacity) is plotted in Figure 8. It is evident from this figure that when the wind power output across all states is combined, the likelihood of extreme events (at the level of aggregate power production) is lowered substantially. This is of particular interest where extremes are due to local instability such as storms and squalls, because the timing and strength of sudden local changes can be difficult to predict in a wind power forecasting system. Under these conditions the confidence interval for a wind farm’s forecast can be expected to increase in size. However from the point of view of aggregate power production the effects of rapid changes are lessened by uncorrelated changes at other wind farms. Another feature of Figure 8 is the apparent high probability of low wind events in NSW. Similarly, in Figure 7 the reliably available capacity value for NSW is very small. These results are likely to be exaggerated due to the high frequency of calms observed at NSW AWS sites. The areas of best wind potential in NSW are located on exposed hilltops whilst AWS sites are located away from such areas. Wind energy sites usually have distributions of wind speed which approximate a Weibull distribution, where the
11
frequency of observations decreases as the wind speed approaches zero. Figure 9 shows the actual distribution of scaled wind speeds at Goulburn AWS and a typical Weibull distribution. It can be seen that the number of calms observed at Goulburn is much greater than would be observed at a typical wind energy site. Further checks showed that the calms tend to be recorded late at night and early morning. The stable atmospheric conditions at night, together with low height above ground (10m) and sheltered location are the likely causes. 14
% Occurrence
12 10 8 6 4 2 0 SA
Vic
NSW
Output95%
Figure 8 Frequency of extreme power production events (extreme low: output95% of installed capacity) 6000 5000
Hours
4000 3000 2000 1000 0 0
4
8
12
16
20
24
28
Wind Speed m/sec Scaled Goulburn Observations
Weibull Fit
Figure 9 Frequency of observations at Goulburn AWS (bars) and a Weibull fit (line) showing a distribution more typical of wind energy sites. Gaps in the distribution are due to the original measurements being taken in integer knots.
12
Duration of extreme power events The average length of low power events was calculated by state to determine the effect of aggregation on the length of NEM-wide low power events. In Figure 10 the average length of low power events is plotted against a threshold value of wind farm capacity level. Figure 10 shows that for capacity values below about 15%, aggregation leads to a decrease in length of time of events crossing below this threshold value. A similar analysis was performed to determine the effect of aggregation on the average length of high power events. Figure 11 shows that for capacity values of greater than about 87% there is a decrease in average length of high power events that have crossed above this threshold in capacity.
Average Time (hours)
8 7 6 5 4 3 2 1 0 0
5
10
15
20
25
30
Max Capacity Level (%)
SA
Vic
NSW
All
Figure 10 Effect of wind farm aggregation on average length of low power events
Average Time (hours)
8 7 6 5 4 3 2 1 0 70
75
80
85
90
95
100
Min Capacity Level (%)
SA
Vic
NSW
All
Figure 11 Effect of wind farm aggregation on average length of high power events
13
Seasonal variation The seasonal variation of power output is plotted in Figure 12. The variation in power output over the calendar year appears to be consistent across the states, with the least wind power being produced in April and May. Aggregation across the states has little effect on the seasonal variation.
Average Capacity %
60.0 50.0 40.0 30.0 20.0 10.0 0.0 1
2
3
4
5
6
7
8
9
10
11
12
Month SA
Vic
NSW
All
Figure 12 Average power by month of year and by state
Wind variability and forecasting Without running a detailed model and validating with real data it is difficult to quantify the expected reduction in forecast error when wind power is aggregated across the south east. However, Focken et al. (2002) conducted a study in Germany using numerical weather prediction and wind farm models with model output statistics. It was found that the standard deviation of the forecast error in combined power output decreased considerably as the wind power was aggregated over a region. The amount of the decrease was strongly dependent on the size of the region. For a region of diameter 730km the reduction was as high as 50% for a 6 hour forecast, becoming less significant for longer time horizons. The reason for the reduction is that forecast errors for individual wind farms tend to be less correlated with each other as the distance between the wind farms increases. Therefore, when combining wind power forecasts over a large region the individual forecast errors partly cancel. Figure 13 shows the relationship between observed changes in power output from hour to hour and their likelihood. When wind power output is combined across the three states, the likelihood of small changes over one hour is increased, and the likelihood of larger changes is reduced. This means that better short-term forecasts of aggregate power output can be expected as wind farms are combined over a large area.
14
% of Installed Capacity -100.0%
-50.0%
0.0%
50.0%
100.0%
Relative Frequency
100.000% 10.000% 1.000% 0.100% 0.010% 0.001% SA
Vic
NSW
All
Figure 13 Hourly incremental change density by state
A forecasting system makes use of both statistical and physical models. Statistical models attempt to predict wind power output based on current and past measurements of wind power and any other variables which may be relevant. Physical models attempt to model the wind using atmospheric physics and upper air measurements. Then model output statistics are commonly used to convert the physical model’s prediction of wind speed and direction to an estimate of wind farm power output. Statistical models perform best when the time horizon of the prediction is short – up to about 5 hours. This is because the wind exhibits a large degree of persistence on short time scales. Physical models are less capable of assessing the current state of the atmosphere, but when the prediction time horizon extends beyond about 5 hours the prediction error is generally lower than for statistical techniques (Giebel et al., 2003). Typically, for a single wind farm the Root Mean Square (RMS) error increases with forecast length to about 15% of installed capacity for a 5 hour ahead forecast using purely statistical techniques. Thereafter the RMS error can be expected to increase more gradually with forecast length as physical weather prediction models are employed. The forecast accuracy of weather prediction models continues to improve with development of those models. Some forecasting models used in Europe are achieving an RMS error of 10-15% of installed capacity for a 36 hour horizon (Giebel et al., 2003). The wind speed prediction from a weather prediction model must be transformed into an estimate of wind farm power. Over time the forecast of wind farm power can be tuned to the measured wind farm output, increasing model predictive reliability. Quantification of forecast uncertainty is another active research field. The uncertainty for a particular forecast varies with the predicted wind speed, because at high wind speeds 15
there is a risk that turbines will reach their cut off limit. Also, forecast uncertainty can increase with atmospheric instability. Knowledge regarding forecast uncertainty is gained by running the models and analysing the statistics of the forecast errors. Prediction of rapid changes in wind can be problematic. The timing and strength of storm fronts can be especially difficult to predict, and such events have the potential to cause the shutting down of wind turbines. With large distribution of wind farms, and turbines within each wind farm, the impact to the network is reduced. However in areas where storms are likely to have substantial impacts to the network it may be possible to employ surface wind measurements or radar as an early warning system. Also, when a large degree of atmospheric instability is evident a lower confidence can be assigned to the wind energy forecast and appropriate mitigating action taken. The accuracy of individual wind farms’ forecasts could be improved if forecasting models incorporated the array of spatial and temporal information provided by a widely distributed network of wind monitoring. The enhanced network of spatial data would be able to be used readily in short term statistical forecasts. This is because the future wind behaviour at a site may be related to the current behaviour at nearby locations. In addition, the use of a large network of data allows more modelling flexibility when predicting the aggregate wind power of a region. An example of an approach to predicting regional wind power is presented in the paper by Nielsen et al. (2002).
16
Analysis of Periodic Components of Wind Power The time series of wind power production, whilst not periodic or stationary, has some cyclical elements which represent physical processes in the atmosphere. Using the Discrete Fourier Transform, a time series can be decomposed into a sum of periodic components having different phases, amplitudes and frequencies. The total variance of the time series is equal to the sum of the variances of the periodic components. It is useful to look at the change in frequency content of power production as the outputs from two or more wind farms are combined. This has implications for the ability to forecast power production. It is known that high frequency variations in power production can be made quite small by combining output from many turbines or several wind farms (Ernst, 1999). This is because high frequency changes tend to be uncorrelated even over relatively short distances. An analysis was performed to study the frequency content of power output for two wind farms as the distance between them increases. This was done in order to understand in greater detail the benefits (and limits to these benefits) from diversification of wind farm areas. The simulated wind farm output from three areas (see Table 3) was combined in turn with that at Sellicks Hill. For each case a Discrete Fourier Transform was calculated. The variance as a function of oscillation period was plotted in a cumulative fashion starting with high frequencies (short period). Since power output was measured as percentage of capacity, the units for variance are percentage squared. It should be noted that for this exercise all wind farm areas were given equal weighting. In other words, the maximum power output from all areas was assumed to be the same. The analysis was conducted over two particular time periods, one representing most of winter and another largely covering summer. The results are shown in Figure 14 and Figure 15. It is evident from these Figures that most of the variance in power is due to longer period events (i.e. lower frequency events) lasting one day or more. There is an overall reduction in variance when the outputs from two wind farms are combined. The reason is that for a large range of frequencies there is a phase difference between the two sites. Figure 14 and Figure 15 also show a marked reduction in slope of the curve as the distance between the sites increases. This effect is most evident for time scales ranging from one to ten days, which represent daily cycles and synoptic-scale events. Certain synoptic-scale events impact upon the whole south east of the continent but there is a time delay between states as the systems move generally from west to east. This effect appears to be stronger in winter than in summer, however this result may be due to some rather unrepresentative data from Bathurst, which like Goulburn showed a tendency to record calms late at night and early in the morning. The simulated power output from Bathurst showed a daily cycle with variance much higher than would normally be expected from a wind energy site. The variance for time scales less than one day was also excessive. Any power variations of period ten days or longer tend to be in phase across the south east. One such variation is the annual seasonal cycle. As discussed above, Figure 12
17
shows that combining the wind power output across the three states doesn’t affect the annual cycle substantially. Table 3 Sites chosen for variance analysis
Location Cape Jaffa Cape Nelson Bathurst
Distance from Sellicks Hill 200km 400km 1000km 1400
(% of capacity)2
1200 1000 800 600 400 200 0 0.01
0.1
1
10
100
Period (days)
Sellicks
Sellicks+Jaffa
Sellicks+CapeNelson
Sellicks+Bathurst
Figure 14 Cumulative variance of power output as wind farm areas are combined. Time period 24 Jun 2002 to 18 Sep 2002 (“Winter”)
1400
(% of capacity)2
1200 1000 800 600 400 200 0 0.01
0.1
1
10
100
Period (days)
Sellicks
Sellicks+Jaffa
Sellicks+CapeNelson
Sellicks+Bathurst
Figure 15 Cumulative variance of power output as wind farm areas are combined. Time period 05 Dec 2001 to 18 Mar 2001 (“Summer”)
18
Examples of Wind Events in South East Australia In winter and early spring, winds are mostly generated by fronts associated with cold air from the Southern Ocean in patterns lasting several days at a time. Figure 16 illustrates the estimated power production during such an event. Broad scale variations are well correlated across the states including relative calms when frontal activity has subsided. The stronger winds reach South Australia first, followed by Victoria then New South Wales. The time lag between SA and NSW may be up to one day or so. Figure 16 shows that aggregation of wind power across the states gives rise to a noticeable reduction in the magnitude (i.e. a smoothing) of high frequency variations. 100 90 80 Capacity %
70 60 50 40 30 20 10 0 07-Aug-01
12-Aug-01
17-Aug-01 SA
22-Aug-01 Vic
27-Aug-01
NSW
01-Sep-01
06-Sep-01
All
Figure 16 Estimated power production - cold fronts during winter
Summer winds, while governed by synoptic patterns, have a stronger diurnal cycle which is related to the heating over the land and the subsequent convective mixing in the atmosphere. Sea breeze effects (due to the difference in temperature between the land and the sea) may be observed. Cold fronts can still cause rapid cooling and a change in wind conditions. Figure 17 shows a diurnal cycle in estimated power production during summer.
19
100
Capacity %
90 80 70 60 50 40 30 20 10 0 07-Feb-03
09-Feb-03
11-Feb-03 SA
Vic
13-Feb-03 NSW
15-Feb-03
All
Figure 17 Estimated power production - diurnal cycle in summer
Figure 18 shows an extended calm event in South Australia whilst moderate winds were present in the other states. Generally there is a period of relative calm between fronts and in this case the calm was longer than normal in South Australia. The long time scales of this pattern indicate that it is caused by the prevailing synoptic conditions over the south east of the continent. Numerical weather prediction models are able to forecast these slow moving changes with a well-defined expected level of accuracy up to about 48 hours ahead. Faster changes are usually forecasted using statistical techniques up to a few hours ahead (Giebel et al., 2003). There is nothing to suggest that this kind of event would present any special difficulty to a forecasting system. 100 90 80 Capacity %
70 60 50 40 30 20 10 0 11-Jun-01
13-Jun-01
15-Jun-01 SA
17-Jun-01 Vic
NSW
19-Jun-01
21-Jun-01
All
Figure 18 Estimated power production – extended calm in South Australia (winter)
20
Figure 19 shows a summer period starting with strong winds in New South Wales and moderate winds in the other states, followed by a strengthening of the wind in South Australia. A diurnal cycle is also evident, as in Figure 17. The changes which take place over a number of days are again due to synoptic conditions over the south east of the continent and these conditions would be taken into account in any forecasting system. The strength of the diurnal cycle in coastal areas is partly influenced by local sea breeze effects which can be difficult to predict. Such phenomena may be better forecasted using statistical techniques up to a few hours ahead.
Capacity %
100 90 80 70 60 50 40 30 20 10 0 30-Jan-03
01-Feb-03
03-Feb-03 SA
Vic
05-Feb-03 NSW
07-Feb-03
09-Feb-03
All
Figure 19 Estimated power production – strong winds in NSW (summer)
21
Wind and NEM Load Data for aggregate load were obtained from NEMMCO covering the period of the wind study. The data were used to look for broad correlations between wind and load. Three wind sites were chosen: Sellicks Hill, Cape Nelson and Bathurst. Attempts were made to fit linear models relating the average daily state load with the wind at the site located in that state. This was performed for winter months and summer months separately. Weekends and the period 25 December to 07 January were excluded. This was so as to exclude causes in load variation which are largely anthropological and not weatherrelated. Load levels during these periods are usually lower regardless of the weather conditions. Two kinds of linear models were used. Firstly the average daily load was related to the average daily wind speed. Secondly, daily average load was related to both the average North/South component and average East/West component of wind speed using a multiple regression. This second model is a way of including wind direction information to see whether it plays an important role. Being a circular or periodic variable, the wind direction itself cannot be used as an explanatory variable in a conventional linear model. Scaled wind speed values were used, with units in ms-1, and the units for load were MWh. Results have been displayed in Table 4, Table 5, Table 6, Table 7 and Table 8. The estimated coefficients are shown, along with the R2 value which is the fraction of variance in NEM load explained by the chosen variable(s). The R2 value provides a measure of the practical significance of the linear model result. However the R2 value below which a result is considered to be of little practical significance is somewhat arbitrary. Here it will be assumed that when R2 is less than 0.10 the finding is of low practical significance. If R2 is between 0.10 and 0.50 the finding is assumed to be of moderate practical significance. In the tables below the results of moderate practical significance are shown in bold. Of the results with moderate practical significance, all of the coefficients were found to be statistically significant at the 0.001 level.
Results The results show that in South Australia daily load is not strongly related to wind speed alone however it is strongly related to the North/South and East/West components of the wind during summer months. The coefficients appear to reflect that high winds from the south or the west bring cooler temperatures and lower load. (Winds from the north tend to be infrequent during summer.) Approximately 47% of the variance in daily average load can be explained in this way. In summer, Victorian load is related to the daily average wind speed at Cape Nelson. Average daily load decreases for higher average wind speeds. This effect explains 21% of the variance in average daily load. When the wind is broken down into its directional components, a linear model explains 28% of the variance in average daily load. In NSW generally, and in winter across all states, the relationship between load and wind was found to be of little practical significance, regardless of statistical significance. This was reflected in the low R2 values (less than 0.10). The wind generation scenario in NSW
22
is distinguished from the other states in that the areas of highest potential (on the Great Dividing Range) are climatically different from the population centres. Table 4 Winter: linear dependence of daily average state load on wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Speed Coefficient 0.7 13.1 -17.0
R2 0.001 0.04 0.03
Table 5 Summer: linear dependence of daily average state load on wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Speed Coefficient -14.6 -118.0 -23.8
R2 0.02 0.21 0.007
Table 6 Winter: linear dependence of daily average state load on northerly and easterly components of wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Northerly Coefficient -1.96 -6.1 -18.6
Easterly Coefficient -3.05 -6.3 -5.3
R2 0.05 0.04 0.05
Table 7 Summer: linear dependence of daily average state load on northerly and easterly components of wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Northerly Coefficient 42.8 54.8 31.5
Easterly Coefficient 31.9 50.9 18.4
R2 0.47 0.28 0.05
The results for South Australia and Victoria in summer warranted some further investigation as to whether the wind in one state could be related to the load in the other state. Table 8 shows that the Northerly and Easterly components of the wind at Sellicks Hill explain 30% of the variance in daily average load in Victoria. Similarly the wind components at Cape Nelson explain 35% of the variance in South Australian average load. Since the peaks in Victorian load are much higher in magnitude, there may be potential for flows of wind power into Victoria from South Australia during summer. There may be particular conditions when this becomes more likely. More sophisticated models would be better able to predict when this might occur.
23
Table 8 South Australia and Victoria, Summer: linear dependence of daily average state load on northerly and easterly components of wind speed in the other state
Sellicks Hill Cape Nelson
Northerly Coefficient 69.5 28.6
Easterly Coefficient 63.4 24.6
R2 0.30 0.35
References Coppin, P.A., Ayotte, K.A., and Steggel, N. (2003), Wind Resource Assessment in Australia – a Planners Guide (available at http://www.csiro.au/weru) Electricity Supply Industry Planning Council (2003), South Australian Wind Power Study Ernst, B. (1999), Short-Term Power Fluctuations of Wind Turbines from the Ancillary Services Viewpoint, AWEA Conference Focken, U. et al. (2002), Short-term prediction of the aggregated power output of wind farms—a statistical analysis of the reduction of the prediction error by spatial smoothing effects, Journal of Wind Engineering and Industrial Aerodynamics, Volume: 90, pp. 231246 Giebel, G. et al. (2003), The State-of-the-Art in Short-Term Prediction of Wind Power – A Literature Overview, Version 1.1, ANEMOS Project, European Commission (available at http://anemos.cma.fr) Nielsen, T. S. et al. (2002), Prediction of regional wind power, Proceedings of Global Wind Power Conference, Paris, 2-5 April 2002
24
100 90 80 Capacity %
70 60 50 40 30 20 10 0 11-Jun-01
13-Jun-01
15-Jun-01
17-Jun-01
19-Jun-01
21-Jun-01
This report was prepared for the Australian Greenhouse Office by the CSIRO Wind Energy Research Unit, October 2003
Published by the Australian Greenhouse Office. © Commonwealth of Australia 2003 ISBN: 1 920840 96 6
This work is copyright. It may be reproduced in whole or in part for study or training purposes subject to the inclusion of an acknowledgment of the source, but not for commercial usage or sale. Reproduction for purposes other than those listed above requires the written permission of the Australian Greenhouse Office. Requests and inquiries concerning reproduction and rights should be addressed to: The Manager Communications Australian Greenhouse Office GPO Box 621 CANBERRA ACT 2601
IMPORTANT NOTICE – PLEASE READ This report has been prepared by the CSIRO for the Australian Greenhouse Office. This report is produced for general information and does not represent a statement of the policy of the Australian Government. While every effort has been made to ensure accuracy and completeness, no guarantee is given, nor responsibility taken by the Australian Government for errors or omissions in the report, and the Australian Government does not accept responsibility in respect of any information or advice given in relation to or as a consequence of anything contained here. The Australian Government and all persons acting for the Australian Government in preparing this report accept no liability for the accuracy of or inferences from the material contained in this publication, or for any action as a result of any person’s or group’s interpretations, deductions, conclusions or actions in relying on this material.
Important CSIRO Disclaimer To the extent permitted by law, CSIRO Atmospheric Research (including its employees and consultants) excludes all liability to any person for any consequences, including but not limited to all losses, damages, costs, expenses and any other compensation, arising directly or indirectly from using this publication (in part or in whole) and any information or material contained in it.
Contents Summary ............................................................................................................................. 1 Study Purpose and Scope .................................................................................................... 3 Introduction ......................................................................................................................... 3 Data Preparation .................................................................................................................. 4 Analysis............................................................................................................................... 7 Analysis of Periodic Components of Wind Power............................................................ 17 Examples of Wind Events in South East Australia ........................................................... 19 Wind and NEM Load ........................................................................................................ 22 References ......................................................................................................................... 24
Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9
Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19
Location of Automatic Weather Stations used in the study- colours represent topographic height in metres ........................................................................... 4 Power curve for Vestas V80-2.0MW,104.0db(A) at sea level....................... 6 Generation duration curves by state ............................................................... 8 Generation duration curves by season............................................................ 9 Generation duration curves by state for hours 8am to 8pm inclusive .......... 10 Generation duration curves by state for hours 9pm to 7am inclusive .......... 10 Reliably available capacity figures by state and times of day...................... 11 Frequency of extreme power production events (extreme low: output95% of installed capacity)............ 12 Frequency of observations at Goulburn AWS (bars) and a Weibull fit (line) showing a distribution more typical of wind energy sites. Gaps in the distribution are due to the original measurements being taken in integer knots. ............................................................................................................. 12 Effect of wind farm aggregation on average length of low power events ... 13 Effect of wind farm aggregation on average length of high power events ... 13 Average power by month of year and by state.............................................. 14 Hourly incremental change density by state.................................................. 15 Cumulative variance of power output as wind farm areas are combined. Time period 24 Jun 2002 to 18 Sep 2002 (“Winter”)............................................. 18 Cumulative variance of power output as wind farm areas are combined. Time period 05 Dec 2001 to 18 Mar 2001 (“Summer”)......................................... 18 Estimated power production - cold fronts during winter .............................. 19 Estimated power production - diurnal cycle in summer................................ 20 Estimated power production – extended calm in South Australia (winter) .. 20 Estimated power production – strong winds in NSW (summer)................... 21
Summary Whilst wind farm development has been accelerated in Australia by the Mandatory Renewable Energy Target (MRET), the majority of proposed wind farms on the mainland to date are located in a restricted number of areas, with particular interest evident in South Australia. This report has been undertaken to investigate the possible benefits of distributing wind farms across the south east of the continent, including both coastal and inland sites in South Australia, Victoria and New South Wales. Using modified Bureau of Meteorology wind data from 9 sites over 4 years, the results of this analysis indicate that more widespread distribution of wind farms (or greater locational ‘diversity’) would result in the following benefits: •
reduction in frequency and average length of extreme power events
•
reduced variation in the average contribution of wind energy
•
reduced volatility of average aggregated output
•
improved ability to forecast total wind power output
•
reduction in variance of total power on time scales up to about ten days long
However it was found that seasonal variations were similar across the south east of the continent and hence large scale aggregation would have little effect on the annual seasonal cycle in wind power output. These results indicate that variability in total wind power output can be reduced to some degree by wider distribution of numerous wind farms but remains substantial because synoptic patterns have a strong influence over the whole region. (Synoptic patterns are broad-scale wind patterns associated with the movement of low and high pressure systems across the continent.) Therefore, as in other countries forecasting would be a valuable tool in managing the impact of the variation. With wind farms and forecasting systems in place it will be possible to tune forecasts to measured wind farm output to improve forecast reliability and to develop appropriate confidence intervals according to the atmospheric conditions. Some correlation was found between the wind and load in the National Electricity Market (NEM) on a state-wise basis. This effect was observed for South Australia and Victoria but only for summer weather patterns. Further analysis would be required to draw useful conclusions on the match between regional load especially peak load and regional wind regimes. A large proportion of variation in load does not appear to be correlated with the variation in wind, and this suggests that forecasting of wind (together with standard load forecasting) will be useful for understanding wind power’s potential to efficiently displace non-renewable generation.
1
In a number of cases, results for New South Wales were limited by the use of Bureau of Meteorology data which were not wholly representative of wind energy sites on the Great Dividing Range. Note: It is important to remember that the results of this study are based on assumed state penetration levels and locational distributions based on current proposals. There has been no technically based analysis to date as to the levels of wind penetration that can be supported by any state. These factors mean that actual state penetration levels may ultimately be quite different from the levels assumed for this analysis. The results of this analysis depend on the assumed distributions and penetration levels and cannot necessarily be extrapolated if the actual distributions and capacities of wind farms in the south east of Australia differ.
2
Study Purpose and Scope The purpose of the work undertaken in this report is to better understand correlations and relationships between wind regimes across a number of states and investigate the effects of aggregating wind power across a wide area in south east Australia. The aspects covered include: •
the statistical availability levels of wind power based on the proportion of installed capacity that is statistically available for at least 95% of the time (referred to here as “reliably available capacity”);
•
the frequency of extreme wind events (lulls and high winds);
•
the duration of extreme wind events;
•
the spatial and temporal properties of the wind; and
•
the relationship with load in the National Electricity Market
Introduction Although wind farm developments to date have tended to be concentrated in a small number of areas, particularly coastal South Australia and Victoria, there are wind farm feasibility studies being undertaken throughout south eastern Australia where good wind resources have been identified. Aggregating wind farm output from diverse locations across the continent provides some ability to smooth out some of the variability in the supply of wind energy from any individual wind farm or single area. The weather systems which are the source of most of this broader scale variability generally move from west to east, taking times of the order of a day or more to traverse the full range of potential generation locations from western South Australia to northern NSW. The systems also evolve as they move and this together with a wide range of local effects allows wind generation in one location to be quite different to that in another. To investigate the extent of this spatial smoothing effect, a number of potential wind generation areas have been selected, the local wind energy generation calculated and statistics of the combined output examined.
3
Data Preparation Wind data were obtained from a number of Bureau of Meteorology Automatic Weather Stations (AWS) sites. The sites were chosen both to obtain a wide distribution across the three States and to broadly represent the areas of greatest current interest to wind farm developers in the south east of the continent, however AWS sites are not generally available for well exposed inland areas. The sites chosen for this analysis are shown in Figure 1 and were: •
South Australia: Coles Point, Sellicks Hill, Cape Jaffa
•
Victoria: Cape Nelson, Kilmore Gap, East Sale
•
New South Wales: Goulburn, Bathurst, Glen Innes
-28
Glen Innes
-30
-32
Bathurst -34
Coles Pt
Goulburn
Sellicks Hill -36
Cape Jaffa -38
Cape Nelson
KilmoreGap East Sale
1300 1200 1100 1000 900 800 700 600 500 400 300 200 100 0
-40
-42 134
136
138
140
142
144
146
148
150
152
154
Figure 1 Location of Automatic Weather Stations used in the study - colours represent topographic height in metres
4
Considering the simultaneous availability of full data sets from all AWS sites chosen, the need to use a whole number of years, and other data limitations, for instance Victorian sites were generally only available since 1999, the time span of the analysis was chosen to be 01 Mar 1999 to 01 Mar 2003, a four-year time interval. Some sites have not recorded half hourly data consistently over that time, swapping to hourly observations. Therefore, it was decided to use hourly observations throughout. Where samples were not taken on the hour the closest available sample was used instead (sometimes the nearest half hourly sample). Times were converted to Eastern Standard Time. Small gaps (up to 4 hours) were filled using linear interpolation. Some larger gaps were filled by combining two neighbouring stations - for example, some poor data at Bathurst spanning about one year were filled using Orange AWS. Also the Coles Point record was rather patchy and was filled using Port Lincoln. Several gaps in the data remained but it was preferred to leave them without further manipulation. This was so as not to artificially degrade or enhance any correlations which exist in the wind data. Overall the simultaneous availability of observations for the cleaned data was about 94%. Gaps were spaced in a fairly random fashion throughout the dataset, however the data availability did vary with calendar month. The lowest was May with 89% and the highest was April with 99%. This could have introduced a small amount of seasonal bias to those results which were not stratified by season, but generally these figures indicate that the data availability was sufficient for this kind of analysis. Wind speeds were scaled to yield overall averages of 8.0 ms-1 for inland sites and 8.5ms-1 for coastal sites as shown in Table 1. This scaling process was applied because 8.0 ms-1 is an approximate figure often regarded as the minimum for economic viability under the current regulatory circumstances, and coastal wind speeds are generally higher than inland (Coppin et al. 2003). Power estimates were obtained by applying the Vestas V802.0MW,104.0db(A) power curve as this represents a typical turbine being chosen for current developments. The turbine curve was interpolated using cubic splines. Appropriate air density figures were used for each site, as shown in Table 1, given its height above sea level. The turbine curve for sea level is shown in Figure 2. The turbine is designed to operate for wind speeds between 4 and 25 ms-1. The real response of a wind farm departs from this curve, firstly because it represents an average for a single turbine, and secondly because wind speeds are not the same across the whole wind farm site. The measured response of a wind farm would resemble the power curve in Figure 2, but would have some scatter and the cutoff at high wind speeds would be smoother. The use of a power curve therefore represents an approximation, and the simulated effect of the cutoff at high wind speed will be more pronounced than in any real system.
5
Table 1 Assumed air density and average wind speed for each site
Site Coles Point Sellicks Hill Cape Jaffa Cape Nelson Kilmore Gap East Sale Goulburn Bathurst Glen Innes
Scaled Average Wind Speed (ms-1) 8.5 8.5 8.5 8.5 8.0 8.5 8.0 8.0 8.0
Air Density (kg/m3) 1.21 1.18 1.21 1.21 1.15 1.21 1.12 1.09 1.09
2500
Power (kW)
2000 1500 1000 500 0 0
5
10
15
20
25
30
Wind Speed (m/sec)
Figure 2 Power curve for Vestas V80-2.0MW,104.0db(A) at sea level
Power output for each site was scaled to reflect an assumed installed capacity for the region as shown in Table 2. This step was intended to give a higher relative weighting to production in Victoria and NSW than existing proposals indicate at this stage in order to explore the impact of increased diversity in wind farm locations across the south east. However South Australian sites were still given the greatest weighting in relative terms reflecting the high number of proposed short-term wind developments and the excellent coastal wind resource. For this analysis, the relative weightings are more important than the absolute capacity levels.
6
Table 2 Weighting of wind farm capacity
SA (1600MW) Vic (1200MW) NSW (800MW)
Site Coles Pt Sellicks Hill Cape Jaffa Cape Nelson Kilmore Gap East Sale Goulburn Bathurst Glen Innes
Capacity MW 600 200 800 600 400 200 400 200 200
7
Analysis
Capacity %
Statistical availability The estimated power output from all regions was combined to analyse the way in which diversification of wind farm areas contributes to the proportion of total installed capacity which is statistically available more than 95% of the time. Cumulative histograms of the power output were calculated and plotted to produce generation duration curves shown in Figure 3 and Figure 4. The curves slope downwards from left to right, since 100% capacity is almost never available, and 0% capacity is almost always exceeded. In the generation duration plots the dotted black line is kept as a reference, being the curve for all sites at all times. 100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA
Vic
NSW
All
Figure 3 Generation duration curves by state
8
Capacity %
100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded Spring
Summer
Autumn
Winter
All
Figure 4 Generation duration curves by season
The results indicate that: •
when the wind power from all three states is combined, there is a tendency towards a flatter generation duration curve with a higher reliably available capacity (Figure 3).
•
the reliably available capacity is generally highest during summer and lowest during autumn (Figure 4).
•
higher reliably available capacity is observed during the daytime hours 8am to 8pm than during the remaining hours of 9pm to 7am which include overnight wind phenomena (Figure 5 and Figure 6).
9
Capacity %
100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA 8am to 8pm
Vic 8am to 8pm
All 8am to 8pm
All
NSW 8am to 8pm
Capacity %
Figure 5 Generation duration curves by state for hours 8am to 8pm inclusive 100 90 80 70 60 50 40 30 20 10 0 0
10
20
30
40
50
60
70
80
90
100
% of time capacity is exceeded SA 9pm to 7am
Vic 9pm to 7am
All 9pm to 7am
All
NSW 9pm to 7am
Figure 6 Generation duration curves by state for hours 9pm to 7am inclusive
Figure 7 shows that when wind power from the three states is combined the reliably available capacity increases considerably. However, the reliably available capacity across all states is less than the sum of the reliably available capacities from each state. It is interesting to compare this statistical quantity with “firm capacity”, a term commonly used to describe the availability levels for traditional fossil fuel generation that is able to be scheduled in the NEM. Wind energy has quite different uncertainty characteristics to
10
fossil fuel generation, which means that wind generation is not well described by the term “firm capacity”. Wind farm output is variable and not always dispatchable, as generation levels at a given time are determined by the variation in the wind resource. However, wind farms are potentially very technically reliable (i.e. little affected by equipment failure) because they contain a large number of relatively small generators. Thus for wind energy, predictability of the wind regime becomes the key network management issue. For these reasons this paper avoids referring to “firm capacity” as its meaning is often misinterpreted when used in relation to the statistical availability levels of wind energy. 14
% Capacity
12 10 8 6 4 2 0 SA
Vic 8am to
NS 9pm to
All All
Figure 7 Reliably available capacity figures by state and times of day
Frequency of extreme power events The frequency of extreme power production events (where an extreme low is defined as output of less than 5% of installed capacity and an extreme high is defined as output of greater than 95% of installed capacity) is plotted in Figure 8. It is evident from this figure that when the wind power output across all states is combined, the likelihood of extreme events (at the level of aggregate power production) is lowered substantially. This is of particular interest where extremes are due to local instability such as storms and squalls, because the timing and strength of sudden local changes can be difficult to predict in a wind power forecasting system. Under these conditions the confidence interval for a wind farm’s forecast can be expected to increase in size. However from the point of view of aggregate power production the effects of rapid changes are lessened by uncorrelated changes at other wind farms. Another feature of Figure 8 is the apparent high probability of low wind events in NSW. Similarly, in Figure 7 the reliably available capacity value for NSW is very small. These results are likely to be exaggerated due to the high frequency of calms observed at NSW AWS sites. The areas of best wind potential in NSW are located on exposed hilltops whilst AWS sites are located away from such areas. Wind energy sites usually have distributions of wind speed which approximate a Weibull distribution, where the
11
frequency of observations decreases as the wind speed approaches zero. Figure 9 shows the actual distribution of scaled wind speeds at Goulburn AWS and a typical Weibull distribution. It can be seen that the number of calms observed at Goulburn is much greater than would be observed at a typical wind energy site. Further checks showed that the calms tend to be recorded late at night and early morning. The stable atmospheric conditions at night, together with low height above ground (10m) and sheltered location are the likely causes. 14
% Occurrence
12 10 8 6 4 2 0 SA
Vic
NSW
Output95%
Figure 8 Frequency of extreme power production events (extreme low: output95% of installed capacity) 6000 5000
Hours
4000 3000 2000 1000 0 0
4
8
12
16
20
24
28
Wind Speed m/sec Scaled Goulburn Observations
Weibull Fit
Figure 9 Frequency of observations at Goulburn AWS (bars) and a Weibull fit (line) showing a distribution more typical of wind energy sites. Gaps in the distribution are due to the original measurements being taken in integer knots.
12
Duration of extreme power events The average length of low power events was calculated by state to determine the effect of aggregation on the length of NEM-wide low power events. In Figure 10 the average length of low power events is plotted against a threshold value of wind farm capacity level. Figure 10 shows that for capacity values below about 15%, aggregation leads to a decrease in length of time of events crossing below this threshold value. A similar analysis was performed to determine the effect of aggregation on the average length of high power events. Figure 11 shows that for capacity values of greater than about 87% there is a decrease in average length of high power events that have crossed above this threshold in capacity.
Average Time (hours)
8 7 6 5 4 3 2 1 0 0
5
10
15
20
25
30
Max Capacity Level (%)
SA
Vic
NSW
All
Figure 10 Effect of wind farm aggregation on average length of low power events
Average Time (hours)
8 7 6 5 4 3 2 1 0 70
75
80
85
90
95
100
Min Capacity Level (%)
SA
Vic
NSW
All
Figure 11 Effect of wind farm aggregation on average length of high power events
13
Seasonal variation The seasonal variation of power output is plotted in Figure 12. The variation in power output over the calendar year appears to be consistent across the states, with the least wind power being produced in April and May. Aggregation across the states has little effect on the seasonal variation.
Average Capacity %
60.0 50.0 40.0 30.0 20.0 10.0 0.0 1
2
3
4
5
6
7
8
9
10
11
12
Month SA
Vic
NSW
All
Figure 12 Average power by month of year and by state
Wind variability and forecasting Without running a detailed model and validating with real data it is difficult to quantify the expected reduction in forecast error when wind power is aggregated across the south east. However, Focken et al. (2002) conducted a study in Germany using numerical weather prediction and wind farm models with model output statistics. It was found that the standard deviation of the forecast error in combined power output decreased considerably as the wind power was aggregated over a region. The amount of the decrease was strongly dependent on the size of the region. For a region of diameter 730km the reduction was as high as 50% for a 6 hour forecast, becoming less significant for longer time horizons. The reason for the reduction is that forecast errors for individual wind farms tend to be less correlated with each other as the distance between the wind farms increases. Therefore, when combining wind power forecasts over a large region the individual forecast errors partly cancel. Figure 13 shows the relationship between observed changes in power output from hour to hour and their likelihood. When wind power output is combined across the three states, the likelihood of small changes over one hour is increased, and the likelihood of larger changes is reduced. This means that better short-term forecasts of aggregate power output can be expected as wind farms are combined over a large area.
14
% of Installed Capacity -100.0%
-50.0%
0.0%
50.0%
100.0%
Relative Frequency
100.000% 10.000% 1.000% 0.100% 0.010% 0.001% SA
Vic
NSW
All
Figure 13 Hourly incremental change density by state
A forecasting system makes use of both statistical and physical models. Statistical models attempt to predict wind power output based on current and past measurements of wind power and any other variables which may be relevant. Physical models attempt to model the wind using atmospheric physics and upper air measurements. Then model output statistics are commonly used to convert the physical model’s prediction of wind speed and direction to an estimate of wind farm power output. Statistical models perform best when the time horizon of the prediction is short – up to about 5 hours. This is because the wind exhibits a large degree of persistence on short time scales. Physical models are less capable of assessing the current state of the atmosphere, but when the prediction time horizon extends beyond about 5 hours the prediction error is generally lower than for statistical techniques (Giebel et al., 2003). Typically, for a single wind farm the Root Mean Square (RMS) error increases with forecast length to about 15% of installed capacity for a 5 hour ahead forecast using purely statistical techniques. Thereafter the RMS error can be expected to increase more gradually with forecast length as physical weather prediction models are employed. The forecast accuracy of weather prediction models continues to improve with development of those models. Some forecasting models used in Europe are achieving an RMS error of 10-15% of installed capacity for a 36 hour horizon (Giebel et al., 2003). The wind speed prediction from a weather prediction model must be transformed into an estimate of wind farm power. Over time the forecast of wind farm power can be tuned to the measured wind farm output, increasing model predictive reliability. Quantification of forecast uncertainty is another active research field. The uncertainty for a particular forecast varies with the predicted wind speed, because at high wind speeds 15
there is a risk that turbines will reach their cut off limit. Also, forecast uncertainty can increase with atmospheric instability. Knowledge regarding forecast uncertainty is gained by running the models and analysing the statistics of the forecast errors. Prediction of rapid changes in wind can be problematic. The timing and strength of storm fronts can be especially difficult to predict, and such events have the potential to cause the shutting down of wind turbines. With large distribution of wind farms, and turbines within each wind farm, the impact to the network is reduced. However in areas where storms are likely to have substantial impacts to the network it may be possible to employ surface wind measurements or radar as an early warning system. Also, when a large degree of atmospheric instability is evident a lower confidence can be assigned to the wind energy forecast and appropriate mitigating action taken. The accuracy of individual wind farms’ forecasts could be improved if forecasting models incorporated the array of spatial and temporal information provided by a widely distributed network of wind monitoring. The enhanced network of spatial data would be able to be used readily in short term statistical forecasts. This is because the future wind behaviour at a site may be related to the current behaviour at nearby locations. In addition, the use of a large network of data allows more modelling flexibility when predicting the aggregate wind power of a region. An example of an approach to predicting regional wind power is presented in the paper by Nielsen et al. (2002).
16
Analysis of Periodic Components of Wind Power The time series of wind power production, whilst not periodic or stationary, has some cyclical elements which represent physical processes in the atmosphere. Using the Discrete Fourier Transform, a time series can be decomposed into a sum of periodic components having different phases, amplitudes and frequencies. The total variance of the time series is equal to the sum of the variances of the periodic components. It is useful to look at the change in frequency content of power production as the outputs from two or more wind farms are combined. This has implications for the ability to forecast power production. It is known that high frequency variations in power production can be made quite small by combining output from many turbines or several wind farms (Ernst, 1999). This is because high frequency changes tend to be uncorrelated even over relatively short distances. An analysis was performed to study the frequency content of power output for two wind farms as the distance between them increases. This was done in order to understand in greater detail the benefits (and limits to these benefits) from diversification of wind farm areas. The simulated wind farm output from three areas (see Table 3) was combined in turn with that at Sellicks Hill. For each case a Discrete Fourier Transform was calculated. The variance as a function of oscillation period was plotted in a cumulative fashion starting with high frequencies (short period). Since power output was measured as percentage of capacity, the units for variance are percentage squared. It should be noted that for this exercise all wind farm areas were given equal weighting. In other words, the maximum power output from all areas was assumed to be the same. The analysis was conducted over two particular time periods, one representing most of winter and another largely covering summer. The results are shown in Figure 14 and Figure 15. It is evident from these Figures that most of the variance in power is due to longer period events (i.e. lower frequency events) lasting one day or more. There is an overall reduction in variance when the outputs from two wind farms are combined. The reason is that for a large range of frequencies there is a phase difference between the two sites. Figure 14 and Figure 15 also show a marked reduction in slope of the curve as the distance between the sites increases. This effect is most evident for time scales ranging from one to ten days, which represent daily cycles and synoptic-scale events. Certain synoptic-scale events impact upon the whole south east of the continent but there is a time delay between states as the systems move generally from west to east. This effect appears to be stronger in winter than in summer, however this result may be due to some rather unrepresentative data from Bathurst, which like Goulburn showed a tendency to record calms late at night and early in the morning. The simulated power output from Bathurst showed a daily cycle with variance much higher than would normally be expected from a wind energy site. The variance for time scales less than one day was also excessive. Any power variations of period ten days or longer tend to be in phase across the south east. One such variation is the annual seasonal cycle. As discussed above, Figure 12
17
shows that combining the wind power output across the three states doesn’t affect the annual cycle substantially. Table 3 Sites chosen for variance analysis
Location Cape Jaffa Cape Nelson Bathurst
Distance from Sellicks Hill 200km 400km 1000km 1400
(% of capacity)2
1200 1000 800 600 400 200 0 0.01
0.1
1
10
100
Period (days)
Sellicks
Sellicks+Jaffa
Sellicks+CapeNelson
Sellicks+Bathurst
Figure 14 Cumulative variance of power output as wind farm areas are combined. Time period 24 Jun 2002 to 18 Sep 2002 (“Winter”)
1400
(% of capacity)2
1200 1000 800 600 400 200 0 0.01
0.1
1
10
100
Period (days)
Sellicks
Sellicks+Jaffa
Sellicks+CapeNelson
Sellicks+Bathurst
Figure 15 Cumulative variance of power output as wind farm areas are combined. Time period 05 Dec 2001 to 18 Mar 2001 (“Summer”)
18
Examples of Wind Events in South East Australia In winter and early spring, winds are mostly generated by fronts associated with cold air from the Southern Ocean in patterns lasting several days at a time. Figure 16 illustrates the estimated power production during such an event. Broad scale variations are well correlated across the states including relative calms when frontal activity has subsided. The stronger winds reach South Australia first, followed by Victoria then New South Wales. The time lag between SA and NSW may be up to one day or so. Figure 16 shows that aggregation of wind power across the states gives rise to a noticeable reduction in the magnitude (i.e. a smoothing) of high frequency variations. 100 90 80 Capacity %
70 60 50 40 30 20 10 0 07-Aug-01
12-Aug-01
17-Aug-01 SA
22-Aug-01 Vic
27-Aug-01
NSW
01-Sep-01
06-Sep-01
All
Figure 16 Estimated power production - cold fronts during winter
Summer winds, while governed by synoptic patterns, have a stronger diurnal cycle which is related to the heating over the land and the subsequent convective mixing in the atmosphere. Sea breeze effects (due to the difference in temperature between the land and the sea) may be observed. Cold fronts can still cause rapid cooling and a change in wind conditions. Figure 17 shows a diurnal cycle in estimated power production during summer.
19
100
Capacity %
90 80 70 60 50 40 30 20 10 0 07-Feb-03
09-Feb-03
11-Feb-03 SA
Vic
13-Feb-03 NSW
15-Feb-03
All
Figure 17 Estimated power production - diurnal cycle in summer
Figure 18 shows an extended calm event in South Australia whilst moderate winds were present in the other states. Generally there is a period of relative calm between fronts and in this case the calm was longer than normal in South Australia. The long time scales of this pattern indicate that it is caused by the prevailing synoptic conditions over the south east of the continent. Numerical weather prediction models are able to forecast these slow moving changes with a well-defined expected level of accuracy up to about 48 hours ahead. Faster changes are usually forecasted using statistical techniques up to a few hours ahead (Giebel et al., 2003). There is nothing to suggest that this kind of event would present any special difficulty to a forecasting system. 100 90 80 Capacity %
70 60 50 40 30 20 10 0 11-Jun-01
13-Jun-01
15-Jun-01 SA
17-Jun-01 Vic
NSW
19-Jun-01
21-Jun-01
All
Figure 18 Estimated power production – extended calm in South Australia (winter)
20
Figure 19 shows a summer period starting with strong winds in New South Wales and moderate winds in the other states, followed by a strengthening of the wind in South Australia. A diurnal cycle is also evident, as in Figure 17. The changes which take place over a number of days are again due to synoptic conditions over the south east of the continent and these conditions would be taken into account in any forecasting system. The strength of the diurnal cycle in coastal areas is partly influenced by local sea breeze effects which can be difficult to predict. Such phenomena may be better forecasted using statistical techniques up to a few hours ahead.
Capacity %
100 90 80 70 60 50 40 30 20 10 0 30-Jan-03
01-Feb-03
03-Feb-03 SA
Vic
05-Feb-03 NSW
07-Feb-03
09-Feb-03
All
Figure 19 Estimated power production – strong winds in NSW (summer)
21
Wind and NEM Load Data for aggregate load were obtained from NEMMCO covering the period of the wind study. The data were used to look for broad correlations between wind and load. Three wind sites were chosen: Sellicks Hill, Cape Nelson and Bathurst. Attempts were made to fit linear models relating the average daily state load with the wind at the site located in that state. This was performed for winter months and summer months separately. Weekends and the period 25 December to 07 January were excluded. This was so as to exclude causes in load variation which are largely anthropological and not weatherrelated. Load levels during these periods are usually lower regardless of the weather conditions. Two kinds of linear models were used. Firstly the average daily load was related to the average daily wind speed. Secondly, daily average load was related to both the average North/South component and average East/West component of wind speed using a multiple regression. This second model is a way of including wind direction information to see whether it plays an important role. Being a circular or periodic variable, the wind direction itself cannot be used as an explanatory variable in a conventional linear model. Scaled wind speed values were used, with units in ms-1, and the units for load were MWh. Results have been displayed in Table 4, Table 5, Table 6, Table 7 and Table 8. The estimated coefficients are shown, along with the R2 value which is the fraction of variance in NEM load explained by the chosen variable(s). The R2 value provides a measure of the practical significance of the linear model result. However the R2 value below which a result is considered to be of little practical significance is somewhat arbitrary. Here it will be assumed that when R2 is less than 0.10 the finding is of low practical significance. If R2 is between 0.10 and 0.50 the finding is assumed to be of moderate practical significance. In the tables below the results of moderate practical significance are shown in bold. Of the results with moderate practical significance, all of the coefficients were found to be statistically significant at the 0.001 level.
Results The results show that in South Australia daily load is not strongly related to wind speed alone however it is strongly related to the North/South and East/West components of the wind during summer months. The coefficients appear to reflect that high winds from the south or the west bring cooler temperatures and lower load. (Winds from the north tend to be infrequent during summer.) Approximately 47% of the variance in daily average load can be explained in this way. In summer, Victorian load is related to the daily average wind speed at Cape Nelson. Average daily load decreases for higher average wind speeds. This effect explains 21% of the variance in average daily load. When the wind is broken down into its directional components, a linear model explains 28% of the variance in average daily load. In NSW generally, and in winter across all states, the relationship between load and wind was found to be of little practical significance, regardless of statistical significance. This was reflected in the low R2 values (less than 0.10). The wind generation scenario in NSW
22
is distinguished from the other states in that the areas of highest potential (on the Great Dividing Range) are climatically different from the population centres. Table 4 Winter: linear dependence of daily average state load on wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Speed Coefficient 0.7 13.1 -17.0
R2 0.001 0.04 0.03
Table 5 Summer: linear dependence of daily average state load on wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Speed Coefficient -14.6 -118.0 -23.8
R2 0.02 0.21 0.007
Table 6 Winter: linear dependence of daily average state load on northerly and easterly components of wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Northerly Coefficient -1.96 -6.1 -18.6
Easterly Coefficient -3.05 -6.3 -5.3
R2 0.05 0.04 0.05
Table 7 Summer: linear dependence of daily average state load on northerly and easterly components of wind speed in that state
Sellicks Hill Cape Nelson Bathurst
Northerly Coefficient 42.8 54.8 31.5
Easterly Coefficient 31.9 50.9 18.4
R2 0.47 0.28 0.05
The results for South Australia and Victoria in summer warranted some further investigation as to whether the wind in one state could be related to the load in the other state. Table 8 shows that the Northerly and Easterly components of the wind at Sellicks Hill explain 30% of the variance in daily average load in Victoria. Similarly the wind components at Cape Nelson explain 35% of the variance in South Australian average load. Since the peaks in Victorian load are much higher in magnitude, there may be potential for flows of wind power into Victoria from South Australia during summer. There may be particular conditions when this becomes more likely. More sophisticated models would be better able to predict when this might occur.
23
Table 8 South Australia and Victoria, Summer: linear dependence of daily average state load on northerly and easterly components of wind speed in the other state
Sellicks Hill Cape Nelson
Northerly Coefficient 69.5 28.6
Easterly Coefficient 63.4 24.6
R2 0.30 0.35
References Coppin, P.A., Ayotte, K.A., and Steggel, N. (2003), Wind Resource Assessment in Australia – a Planners Guide (available at http://www.csiro.au/weru) Electricity Supply Industry Planning Council (2003), South Australian Wind Power Study Ernst, B. (1999), Short-Term Power Fluctuations of Wind Turbines from the Ancillary Services Viewpoint, AWEA Conference Focken, U. et al. (2002), Short-term prediction of the aggregated power output of wind farms—a statistical analysis of the reduction of the prediction error by spatial smoothing effects, Journal of Wind Engineering and Industrial Aerodynamics, Volume: 90, pp. 231246 Giebel, G. et al. (2003), The State-of-the-Art in Short-Term Prediction of Wind Power – A Literature Overview, Version 1.1, ANEMOS Project, European Commission (available at http://anemos.cma.fr) Nielsen, T. S. et al. (2002), Prediction of regional wind power, Proceedings of Global Wind Power Conference, Paris, 2-5 April 2002
24
