Master's Programme in Energy Smart Innovation in
the Built Environment, 120 credits
Analysis of AEP prediction against
production data of commercial wind
turbines in Sweden
Master Thesis in Construction
Engineering with specialization in
Renewable Energy, 30 credits
Halmstad, 2021-05-31
Aromal Sugathan, Sean Gregory
HALMSTAD
UNIVERSITY
Halmstad University • PO Box 823 • SE-301 18 Halmstad • Sweden Phone +46 35 16 71 00 • registrator@hh.se • CIN 202100-3203
Analysis of AEP prediction against production data of
commercial wind turbines in Sweden
Master Thesis in Construction Engineering with specialization in Renewable Energy
By
Aromal Sugathan: arosug19@student.hh.se Sean Gregory: seagre19@student.hh.se
Program: Masters Programme in Energy Smart Innovation in the Built Environment Supervisor: Dr. Erik Möllerström
Examiner: Dr. Mohsen Soleimani Mohseni Halmstad University, Sweden
May 2021
i
Abstract
Based on data from 2083 wind turbines installed in Sweden since 1988, the annual energy production (AEP) predictions considered at the project planning phases of the wind turbines in Sweden have been compared to the wind-index-corrected production data. The production data and the predicted AEP data are taken from Vindstat, a database that collects information directly from wind turbine owners in Sweden. The mean error for all analyzed wind turbines was 11.9%, which means that, overall, the predicted AEP has been overestimated. There has been improved accuracy with time and error in prediction decreasing from 12% to 6.3% for wind turbines installed in the 2000s and 2010s, respectively. However, the overall improvement in accuracy seems to have stagnated around 2005 despite the refinement of forecasting methods and better data availability.
From the results analyzed for effects of terrain, the error is smaller for wind turbines in forest areas than in open terrain, indicating that the complexity of forest terrain is not the reason behind the error. Also, there is no apparent increase of error with wind farm size, which could have been expected if the wind farm blockage effect was a primary reason for the overestimations.
Comparison between significant wind turbine manufacturers Vestas and Enercon in the Swedish context, the error was more prominent for Enercon.
Key Words – Energy assessment, validation, wind power, Sweden, P50, AEP, WCP
ii
Acknowledgments
This study was motivated by the study previously done by (Möllerström and Lindholm, 2020). This study considered a more extensive data set than the previous study and analysis on more factors.
We owe sincere thanks and gratitude to our supervisor Dr. Erik Möllerström for motivating us to continue his previous work ‘Evaluation of AEP Predictions for Commercial Wind Farms in Sweden’ (Möllerström and Lindholm, 2020). His support and guidance in this thesis helped us to complete it within the planned time frame. We also thank Halmstad University for its computer lab facility, which enabled us in the data collection for the thesis.
Furthermore, the authors thank all our professors in the program “Energy Smart Innovation in the Built Environment” at Halmstad University for their continued support and guidance, enabling us to do this study and report.
Aromal Sugathan, Sean Gregory
iii
Table of Contents
1 Introduction ... 1
1.1 History and background ... 1
1.1.1 Wind power in Sweden ... 2
1.2 What causes wind and hence wind power ... 5
1.3 Wind Energy Forecast ... 5
1.3.1 Types of wind energy forecast based on time horizons ... 5
1.3.2 Types of wind energy forecast based on the applied methodology ... 6
1.3.3 Wind Resource Assessment (WRA) using the WindPRO and WAsP software ... 7
2 Literature Study ... 8
2.1 Major factors affecting AEP prediction of wind turbines ... 8
2.1.1 Wake Effect in Wind Farms ... 8
2.1.2 Type of terrain ... 8
2.2 Improvement of AEP of wind farms with time ... 10
2.2.1 Before the 1990s ... 10
2.2.2 The 1990s ... 11
2.2.3 Since the 2000s ... 12
2.3 Cases of overestimation of AEP predictions ... 13
3 Research Question ... 15
4 Methodology ... 16
4.1 Data and Software ... 16
4.1.1 ArcGIS ... 18
4.2 Analysis Method ... 19
4.2.1 Normalizing of production data... 19
4.3 Wind Index Correction ... 20
4.4 Data Exclusion ... 22
5 Results ... 23
6 Discussion ... 31
7 Conclusions ... 33
References ... 34
Appendix-I – Wind Turbines used in the analysis ... 39
iv
List of Figures
Figure 1-1 The world annual addition and previous year’s capacity of wind power ... 3
Figure 1-2 Top 10 countries in the world to have the most wind power installed as in 2018 ... 3
Figure 1-3 Size comparison of wind turbines installed in Sweden over time ... 4
Figure 4-1 Locations of wind turbines (green). Locations of index sites used for typical wind-year for normalization (red) ... 17
Figure 4-2 ArcGIS Mapping Software Interface ... 18
Figure 4-3 Wind Turbine Location, Index Location and Terrain as seen in ArcMap Software ... 19
Figure 4-4 Yearly mean correlation index of all 77 sites ... 21
Figure 4-5 All monthly production data of a single wind turbine normalized for availability (Ea) is plotted against the correlation index for the same month with the index location closest to the same wind turbine (In) ... 22
Figure 5-1 Error of P50-evaluations compared to the production-based WCP depending on the installation year plotted with the standard deviation (blue) and the number of wind turbines (red). ... 23
Figure 5-2 The time development of the AEP prediction accuracy divided on terrain type, manufacturer and wind farm size. ... 26
Figure 5-3 The development of the AEP prediction accuracy with time (blue) divided on terrain type with the number of wind turbines for each year (red). ... 27
Figure 5-4 The development of the AEP prediction accuracy with time (blue) divided among the manufacturers with the number of wind turbines installed for each year (red). ... 28
Figure 5-5 The development of the AEP prediction accuracy with time (blue) divided on wind farm size with the number of wind turbines installed for each year (red). ... 29
Figure 5-6 Mean error for wind farms of different size ... 30
List of Tables
Table 1-1 Prediction Horizon in Wind Power Forecasting ... 6Table 4-1 Wind Turbine Data from Vindstat. Data excluded as per criteria in section 4.4 ... 16
Table 5-1 Results ... 24
v
Abbreviations
AE – Auto Encoder
AEP – Annual Energy Production ANN – Artificial Neural Network AR – Auto Regressive model
ARX – Auto Regressive Exogenous model CNN- Convolutional Nueral Network DTU – Technical University of Denmark ELM – Extreme Learning Machine FLM– Fuzzy Logic Model
HIRLAM - High-Resolution Limited Area weather prediction Model MOS – Model Output Statistics
MLP – Multi-Layer Perceptron algorithm
MODA – Multidimensional Optical Depth Algorithm NWP – Numerical Weather Prediction
RBFNN – Radial Basis Functional Neural Network algorithm RNN– Recurrent Nueral Network
ROI– Return Of Investment SKM–Sparse Kernel Machine
WPPT – Wind Power Prediction Tool
WCP – Wind Index Corrected annual production WRA – Wind Resource Assessment
WTG – Wind Turbine Generator
1
1 Introduction
Wind energy is a renewable energy source characterized by the lowest cost of electricity production and the largest resource available. Therefore, many countries are moving to wind power as a substantial opportunity for future power generation (Wang et al., 2011). To achieve a clean environment, reducing the dependency on fossil fuels and increased integration of renewable- based energy sources is vital. Wind power has become the national policy to combat climate change for many countries (Hanifi et al., 2020). Effective energy production planning is vital for energy companies, especially in the area of forecasting. As the electrical grid evolves, planning and procedures must adapt to these changes. Accurate forecasting techniques are essential to predict wind power generation to make it more economical. For the successful implementation of wind farms, an accurate forecast of wind energy is essential and crucial in the planning process (Nazir et al., 2020). Wind energy prediction also plays an integral part in the feasibility study of the location of wind farms. Most importantly, the modeling of the expected financial returns over the lifetime of the wind farm project can be achieved from a proper AEP prediction. In order to confirm the accuracy of the methodologies and assumptions made at the pre-construction stage, there is a need to periodically check the results of historic predictions against actual production achieved once projects are operational (DNV GL, 2019).
1.1 History and background
Wind energy made its first steps centuries ago with vertical axis wind turbines found in Persian- Afghan borders around 200 BC and horizontal axis wind turbines in Netherlands and Mediterranean much later (Möllerström et al., 2019). In the USA during the 19th century, 6 million small wind turbines were used to pump the water. James Blyth built the first electricity-generating wind turbine in Scotland in 1887(Gipe, 2018). The first large wind turbine to generate electricity was installed in Cleveland, Ohio, in 1888 (Kaldellis and Zafirakis, 2011). During the end stage of World war I, 25-kW wind turbines were widespread throughout Denmark (Kaldellis and Zafirakis, 2011).
Meanwhile, active research on large-scale wind turbines was gaining speed in Europe. The governments of the US, Canada, Sweden, Germany and many other countries focused on funding large-scale prototypes that, in the end, did not develop into commercial turbines (Möllerström, 2019). In contrast, in Denmark, the government supported the market for small wind turbines, so that Danish farmers bought Danish-built turbines (10-30 kW). Thus the Danish wind turbine market started growing and gained widespread attention. This resulted in several Danish manufacturers ready to export wind turbines to California when a market started there around 1980 due to combined federal and state subsidies. These Danish turbines evolved into today’s commercial products (Gipe, 2018). Denmark took the leadership role in wind power development and took a path different from the USA. Unlike American wind turbines in the hands of the aerospace industry, Danish technology grew up in the hands of the agricultural sector. During the late stages of world war-I, more than one-fourth of rural power stations used wind power, with the widespread use of 25 kW machines throughout Denmark. In Germany, Ulrich Hütter later became famous for designing slender and fast rotating two-bladed downwind turbines experimented with different designs, including those that drove an asynchronous generator directly
2
coupled to an electric utility network. The availability of inexpensive oil from the middle east reduced the program’s momentum during the 1950s (Pasqualetti et al., 2004).
1.1.1 Wind power in Sweden
In 1956 an investigation of future energy supply in Sweden was conducted, and the potential for wind power was acknowledged. A 1.2 MW test wind turbine was initially proposed but never materialized for the next 20 years (Möllerström, 2019). After the oil crises in the 1970s, Swedish energy policy began to promote the systematic use of energy and the diffusion of renewable energy resources, mainly to reduce the dependence on imported oil (Söderholm et al., 2007).
In 1974, the state wind power program was launched to decrease the country’s dependence on oil.
In 1977, the Kalkugnen, an experimental 60 kW turbine, was constructed and erected in the northern Uppland coast near Älvkarleby by the National Swedish Board for Energy Source and Saab-Scania AB. However, the turbine later collapsed due to an accident involving the replacement of rotor blades. The wind data from the Kalkugnen wind turbine was used to test early wind turbine wake models. After erecting the Kalkugnen, The National Swedish Board for Energy Source suggested constructing one to three large wind turbine prototypes from different manufacturers. The Swedish government suggested building two prototypes with turbines of 70 to 90 m in diameter, a tower height equal or above the diameter, and a rated power of 2 to 4 MW, having 2 to 3 turbine blades. In total, five bids were made, resulting in the turbines at Maglarp and Näsudden The Maglarp was in operation for 11 years and produced 37 GWh. The Näsudden has been in operation from 1984 to 1988, and the turbine was dismantled in 1991. The tower was then used for Näsudden II. It was in operation from 1993 to 2006 and was decommissioned due to a gearbox failure (Möllerström, 2019).
The first phase of the Swedish wind energy research program ended in 1985. During the latter half of the 1980s, wind power development shifted toward wind turbines from foreign manufacturers.
In 1991 through the introduction of investment subsidies, the number of actors increased, and predominantly Danish and German wind turbines were installed in Sweden’s growing wind power market (Åstrand and Neij, 2006). Apart from Sweden, Denmark also experienced oil crises in the 1970s; unlike Denmark, however, this resulted in investments in nuclear power, and the Swedish wind project developments were very slow until the 2000s (Enevoldsen, 2016).
Today there is no commercially successful wind turbine manufacturer in Sweden. However Swedish firms like ABB and SKF is a major component supplier for the wind power industry (Möllerström, 2019). As of 2017, wind energy penetration levels continued to rise, led by Denmark with a 40% use of its electricity from wind energy, Uruguay, Portugal and Ireland with over 20%;
Germany with 16%; Spain and Cyprus with about 20%; and the major markets of China, the US and Canada with 4%, 5.5% and 6% energy coming from wind energy respectively (Barbosa de Alencar et al., 2017).
Global wind power generation has reached about 20% annually within the past decade. Figure 1-1 (image used with the permission of (Nazir et al., 2020)) depicts the annual addition and previous year’s capacity atlas of wind power from 2014 to 2018. The top 10 countries to have the most wind power installed in 2018 is seen in Figure 1-2 (image used with the permission of (Nazir et al., 2020)).
3
Figure 1-1 The world annual addition and previous year’s capacity of wind power (Nazir et al., 2020)
Figure 1-2 Top 10 countries in the world to have the most wind power installed as in 2018 (Nazir et al., 2020)
As the installed capacity of wind turbines increases dramatically, the requirements for solving various problems become more challenging. These problems include appropriate market design,
4
real-time grid operations, ancillary service requirements and costs, electricity market clearing, competitive power quality, transmission capacity upgrades, power system stability and reliability, and interconnection standards (Breslow and Sailor, 2002) (Lund, 2005). Effective wind forecasting is one of the most efficient ways to overcome many of these problems. Forecasting tools can enhance the position of wind by dealing with the intermittence nature of wind. The impact of the cost of wind energy can be reduced to a significant extent if the wind energy can be scheduled by making use of accurate wind forecasting. Therefore, the improvement of wind power and wind speed forecasting tools has a significant technical and economic impact on the system, increasing wind power penetration (Wang et al., 2011).
Figure 1-3 Size comparison of wind turbines installed in Sweden over time
Figure 1-3 shows the overall increase in the size of the wind turbines installed in Sweden over the years. During the early stages, turbines were generally smaller in capacity with low rotor diameter and low hub height. Theoretically, when the hub height of wind turbines increases, access to good wind speed increases. Likewise, a larger rotor diameter means that the area absorbing wind power increases, increasing the turbine’s power output (Letcher, 2017). The effect of surface roughness in wind speed diminishes with height, so wind speed increment occurs within a mixed layer and mitigates as the height approaches the boundary layer height. Since the overall cost increases with the increase in hub height, the optimal hub height ensures the optimum capital expenditure (Lee et al., 2019). Manufactures like Vestas, Enercon initially manufactured turbines of rated power less than 500-kW (Vestas V20-100, V25-200 and Enercon E30). Later, they started manufacturing turbines above 1MW (Vestas V66-15MW, Enercon E66) of output and introduce them to the Swedish market in the next 15 years. So the trend is that the output power of turbines increases with time, and the hub height and rotor diameter also increases. The infrastructure and investment in the wind power industry keep on increasing, and thus the accurate AEP predictions become crucial.
5 1.2 What causes wind and hence wind power
Wind power generation occurs by the contact of the wind with the blades of the wind turbine.
Wind causes the blades to rotate about a common axis due to the nature of the blade design.
Rotation of the blades converts wind speed into mechanical energy that drives the rotor of the wind generator, which produces electricity. According to (Dutra et al., 2011), tropical regions are warmer than polar regions. Consequently, the warm air found in the low altitudes of the tropical regions rises since warm air is lighter than cold air. The warm air is then replaced by a mass of cooler air that displaces from the polar regions. The displacement of air masses is caused by atmospheric pressure differences between two regions and are influenced by natural effects such as continentally, latitude, altitude, sea level, and soil roughness. Accurate wind forecasting is critical to have a reliable power system. However, the unstable intermittent and nature of the wind speed make it very difficult to predict it accurately (Wang et al., 2017).
The electricity produced from wind generation is obtained from the wind’s kinetic energy. The kinetic energy gets converted into mechanical energy by converting the wind force into a torque that acts on the rotor blades causing it to rotate. The amount of energy generated by wind is a function of its speed and mass and can be calculated by the kinetic energy equation (Patel, 2005).
The power available in the wind cannot be fully utilized by the wind turbine for the generation of electricity. According to (Dutra et al., 2011), an index called power coefficient Cp must be introduced to consider this limitation. Cp can be defined as the fraction of the available wind power that the rotor blades can extract. According to Betz’s law, a wind turbine cannot convert more than 59.3% of the wind’s kinetic energy into mechanical energy at the rotor (Cp ≤ 59.3%).
It means only 59.3% of the energy contained in the airflow can theoretically be extracted by a wind turbine (Thomas and Cheriyan, 2012). The potential of electric energy produced from wind generation is obtained by converting the wind’s kinetic energy into mechanical energy (Barbosa de Alencar et al., 2017).
1.3 Wind Energy Forecast
The thin line between economic loss and economic gain in wind power investments is partly dependent on how good electricity production is estimated before the wind farm’s construction.
One primary use for wind power forecasting is to reduce the risk of uncertainties in the wind, allowing higher diffusion. It is also vital for maintenance planning and the determination of required operating equipment. Wind power forecasting is classified based on either time horizons or applied methodology of forecast (Hanifi et al., 2020).
1.3.1 Types of wind energy forecast based on time horizons
There exist several methods for predicting wind power. Predictive horizons can be divided into four major time scales depending on different functional requirements, summarized in Table 1-1 (Hanifi et al., 2020).
6
Table 1-1 Prediction Horizon in Wind Power Forecasting (Hanifi et al., 2020)
1.3.2 Types of wind energy forecast based on the applied methodology
The evaluation of wind potential in a region requires systematic data collection and analysis on wind speed and regime. A rigorous assessment requires specific surveys of the region where the wind farm will be placed. Wind forecasting methods can be broadly classified into the following three categories: (i) physical methods; (ii) statistical methods; and (iii) hybrid methods (Gasch and Twele, 2011; Manwell et al., 2010).
(i) Physical Methods
Physical methods use detailed physical characterization to model wind turbines/farms. This modeling effort was carried out by downscaling the data from numerical weather prediction (NWP), which requires describing the area, such as roughness, obstacles, and weather forecasting data of temperature and pressure. In order to determine wind speed, these variables are used in complex mathematical models that are time-consuming. Then, the predicted wind speed will be transferred to the related wind turbine power curve to forecast wind power. This method thus depends on physical data (Jung and Broadwater, 2014). This method is suitable for medium to long-term wind power prediction, although it is computationally complex and needs considerable computing resources (Zhang et al., 2019).
(ii) Statistical Methods
The statistical method is based on developing the non-linear and the linear relationships between NWPs data (such as wind speed, wind direction and temperature) and the generated power.
Previous history data will be used as the training data to define the statistical relationship. The model is then fine-tuned by comparing the model prediction with the online measured power. The model is then ready to predict the NWP forecast for the next few hours. This method is inexpensive and easy to model. It is for short-term periods, and as the estimation time increases, its prediction accuracy decreases. Statistical methods can be further divided into time series models and artificial neural networks (ANN). Time series models apply historical data to generate a mathematical model for developing the model, estimating parameters and checking simulation characteristics. Examples of this kind of model are ARX and AR models. ANNs can identify the non-linear relationships between input features and output data. One reason for the tendency to use neural networks is to avoid the complexity of the mechanical structure in wind turbines. An ANN model contains an input layer, one or more hidden layers, and an output layer, where the historical data are fed for training and testing. Some examples of this kind of model use algorithms such as MLP and RBFNN (Hanifi et al., 2020).
7 (iii) Hybrid Methods
The different forecasting methods, such as fuzzy logic models and ANNs, are combined to form hybrid methods (Hong and Rioflorido, 2019). This method aims to retain the advantages of each method and improve the overall accuracy. Diverse predictive models are often developed using multiple algorithms and different training datasets in statistics and machine learning. This process is called ensemble modeling. It is a more advanced type of hybrid forecasting method. A hybrid method may not always lead to a better result but has fewer risks in most situations (Jung and Broadwater, 2014). Some examples of hybrid models are ELM optimized by MODA, Elman and MLP Network and Wavelet decomposition and Elman network (Hanifi et al., 2020).
Physical methods are significantly complex and need considerable computing resources but are suitable for medium to long-term prediction. On the other hand, statistical methods performed better in short-to-medium-term periods that were easily modeled and inexpensive. Combining these two methods with their merits gives rise to hybrid methods (Hanifi et al., 2020).
1.3.3 Wind Resource Assessment (WRA) using the WindPRO and WAsP software
Wind resource assessment is the procedure of estimating the wind power potential at one or several sites or over a region. The result of wind resource assessment is an estimate of the mean wind climate at one or several sites, in the following forms:
• Wind direction probability distribution - the frequency distribution of wind directions at the site.
• Sector-wise wind speed probability distribution functions - the frequency distributions of wind speeds at the site (Mortensen, 2016).
Wind resource assessment provides essential inputs for the siting, sizing and detailed design of the wind farm. windPRO is a wind modeling software created by EMD to design and plan single wind turbines and wind farms. The WAsP is another leading software that provides these results crucial in planning a wind turbine or many wind turbines that constitute a wind farm. A site assessment is usually carried out for the siting of individual wind turbines. This will provide estimates for each wind turbine site of the 50-year extreme wind, shear of the vertical wind profile, flow and terrain inclination angles, free-stream turbulence, wind speed probability distribution and added wake turbulence (Mortensen, 2016). windPRO is used to import all data into the program, and WAsP is the solver and is in charge of wind or energy simulation and generation of resource maps. WAsP uses the linear atmospheric model to extrapolate wind climate data within a region by considering orography and roughness. This model uses the linear components of Navier-Stokes equations to solve wind speed at different locations (Acker and Chime, 2011).
8
2 Literature Study
2.1 Major factors affecting AEP prediction of wind turbines
The largest wind turbines in production today include multi-megawatt machines with rotor diameters spanning over 100 meters. The power that a wind turbine can extract is proportional to the cube of the inflow velocity of air seen by the turbine. It means, if the wind velocity is doubled, the power output from the turbine will be eight times higher than before. While planning new wind turbine installations, it is crucial to find locations where the wind speed is high since this yields a higher electrical energy output. The major factors affecting the AEP prediction of wind turbines and wind farms are discussed below.
2.1.1 Wake Effect in Wind Farms
The wake behind the turbines often disturbs the operation of wind turbines downstream. The velocity in the turbines’ wake is lower than in the approaching wind. Therefore, the downwind turbines exposed to the wakes will produce less electrical energy than the upstream ones.
Additionally, the turbulent nature and the velocity shear of the wind turbine wake increases the fatigue loads on the downstream turbines, which reduces their operational lifetime. Because of these reasons, it is essential to carefully plan the wind farm layout in order to find optimal turbine locations. In offshore areas, the wind is comparatively undisturbed and typically more substantial than over land, making offshore areas suitable for building new wind farms. However, offshore wind farms have the disadvantage that their construction and maintenance costs are significantly higher than onshore wind farms (Hyvarinen, 2018).
Shortage of suitable offshore regions with sufficiently shallow waters for building wind farms can also promote onshore wind-turbine installations. Land-based wind power is, for these reasons, more common than offshore wind power. With long distances of undisturbed winds, large plains are ideal when building wind farms over land. However, onshore wind turbines are often built over remote mountainous regions, where high wind velocities can still be found and where the turbines do not disturb nearby living human residents (Hyvarinen, 2018).
The interaction of the flow between wind turbines in a wind farm results in higher fatigue loads and lower power output on the downstream turbines than the upstream turbines experiencing the free-stream conditions. Depending on the arrangement of wind turbines in a wind farm and the distance between the turbines, the power generation could drop up to 40% when the wind turbine operates within an array other than with free-stream flow (Corten et al., 2004). Moreover, the increase of fatigue loads happening on the downstream turbines due to wake interference can be up to 80% and considerably shorten the rotor’s lifetime (Sanderse, 2009).
2.1.2 Type of terrain
Each wind turbine has a unique power curve, showing the overall performance of the wind turbine, even when the detailed components of the wind turbine generating system are unavailable. By discovering more about the wind, its behavior and how it flows across the ground, we can use more of its resources for wind-energy applications and extract more energy. The wind conditions that a wind turbine perceives depend on many parameters, including the local climate and the environment where the turbine is installed. If a wind turbine is built in a mountainous area, then
9
the local wind conditions will significantly influence the turbine performance. Over mountains, the winds are generally strong since the flow is accelerated over the upwind mountain slopes (Hyvarinen, 2018).
Meanwhile, valley winds are often influenced by the wind direction at the site. The flow can sometimes be accelerated when the wind blows along the valley’s length, while a large wake may form on the leeward side of a mountain in other situations, which significantly reduces the local wind speed. Nearby mountains can also influence the wind conditions seen over hills and ridges in their vicinity. Therefore, flow around onshore turbines is in many cases highly dependent on the surrounding terrain features. The wake of a wind turbine built over a hilly or mountainous region may be disturbed by the flow accelerations and retardations introduced by the local terrain features. Therefore, a careful selection of the wind turbine placements is even more critical at wind sites with significant topographical variations (Hyvarinen, 2018).
Around 70% of Sweden’s land area is covered by forest. Wind mapping efforts (Bergström et al., 2013) have shown good wind resources in Swedish forests. Apart from acquiring land at lower prices, wind turbines installed in forest areas can be deployed far away from residential areas, reducing the negative aspects of wind turbines such as noise emissions and visual impacts. During the last couple of years, wind power in Sweden was growing considerably, most of all perhaps in Swedish forests. The wind field above forests is often characterized by high turbulence, increased wind force with height (vertical wind shear), and strong wind veer with height (vertical wind veer).
Optimum hub heights are not only based on financials and other constraints, but winds have to comply with the IEC standard for turbulence (class A, B or C). The intensity of turbulence decreases with height, even above forests, proving higher hub heights to be a good solution. Due to the cubic effect, more turbulence means more available kinetic energy for a given mean wind speed. However, traditional horizontal axis wind turbines typically will not capitalize on this and decrease performance with turbulence. Due to being omnidirectional and thus insensitive to wind direction changes, vertical axis wind turbines have shown better performance with higher turbulence (Möllerström et al., 2016). Developers have shown concern for big loads and underperformance when building wind farms on forestall terrain. However, even though knowledge is rapidly growing, there is still not much information about turbulence over forests.
(Mohr et al., 2018).
Social acceptance is also an essential factor in establishing wind turbines in ecologically sensitive areas like forests. The social opposition is a risk to be considered in the construction phase, but as the wind project comes into operation, the social opposition decreases (Devine‐Wright, 2005).
According to (Khan, 2003) the development of Swedish wind projects in a region is supported by public acceptance of wind power in the local communities. The local community must approve wind projects in Sweden (Pettersson et al., 2010), and, unlike Denmark, the municipality does not have to allocate areas for wind projects. However, around 28% of Swedes have a negative attitude towards wind power (Pedersen and Waye, 2002). Those who oppose wind projects have higher incomes (Ek, 2005). Opposition to wind turbines is strongest to those positioned in the landscape, as surveys have revealed. The strongest opposition against wind turbines is against those positioned on mountains and forests (Söderholm et al., 2007). The Swedes would prefer higher energy costs to have wind turbines in such locations (Ek and Persson, 2014).
10
Wind energy has been utilized onshore for power generation for more than two thousand years.
However, the history of offshore wind power generation is relatively recent. In recent times, the wind power sector has begun to move offshore, i.e., to use both space and good wind speeds on the open sea for large-scale electricity generation. Offshore wind farms are installed on the continental shelf area about 10 km away from the coast and 10 m in depth. Compared to the turbines installed on land, offshore wind turbines should be fixed on the seabed, demanding a more solid supporting structure. Submarine cables are required to transmit electricity, and special vessels and equipment are needed for building and maintenance work. These factors make offshore turbines double or triple the cost of onshore wind turbines (Zhixin et al., 2009).
Offshore wind turbines are less obtrusive than turbines constructed on land, as their apparent size and noise can be mitigated by distance. Since water has less surface roughness than land, the average wind speed is substantially higher than open water. Capacity factors of offshore wind turbines are higher than onshore and nearshore locations, allowing offshore turbines to utilize shorter towers, making them less visible. In addition, installing wind turbines offshore has many advantages over onshore development. On land, there are difficulties in transporting large components and opposition due to various siting issues, such as noise and visual impacts, limiting the number of suitable locations for wind farms. Offshore locations can benefit from the high capacity of marine shipping and handling equipment, which far exceeds the lifting requirements of multi-megawatt wind turbines. Large wind farms tend to be on relatively remote areas on land, so electricity should be transmitted over long power lines to cities. Offshore wind farms are usually closer to coastal cities and require relatively shorter transmission lines, yet far enough to reduce visual and noise impacts (Bilgili, 2011). Also, due to the better quality of the wind resource in the sea, where wind speed is usually increasing with the distance to the coast, and fewer turbulence effects, the lesser fatigue increases the lifetime of the offshore wind turbine generator (Esteban et al., 2011).
2.2 Improvement of AEP of wind farms with time
According to a review on wind power short-term prediction (Costa et al., 2008), the development in energy prediction for wind power can be classified into 3 phases, (i) before the 1990s, (ii) the 1990s and (iii) since the 2000s.
2.2.1 Before the 1990s
During the 1980s, many important works were published. Notis et al. developed a method to predict wind speed 24 h ahead with a time-step of one hour, intended for load scheduling (Notis et al., 1983). Geerts developed ARMA models and a Kalman filter oriented to the integration of wind energy into the grid. It predicted wind speed with a forecast horizon of 24 h and an hourly time-step (Geerts, 1984). McCarthy made wind predictions using a pocket programmable calculator HP41CX for some wind farms in California from 1985 to 1987. The program was based on local upper-air observations and meteorological observations. It had a forecast horizon of up to 24 h. It outperformed the persistence method and climatology for daily average wind speed (Giebel, 2004).
Kaminsky et al. based their method on a definition of different synoptic weather categories. They worked with a time-step of 15 min to predict wind speed through a regression over the 90 past
11
steps. They concluded that the different synoptic weather categories required different regression methods (Kaminsky et al., 1985). Bossanyi published results for ARMA models forecasting wind speed with time-steps of 2 sec and 1 min. This model had a forecast horizon of up to 10 steps (Bossanyi, 1985). Bailey and Stewart pointed out the requirement for databases at a local level and short time intervals. They proposed evaluating existing databases such as meteorological monitoring networks, synoptic analysis products, forecast products, climatological analyses and digitized orography (Bailey and Stewart, 1987). Before the 1990s, the forecasting methods were generally limited to a time horizon of 24h and limited availability of wind data.
2.2.2 The 1990s
Troen and Landberg proposed a study in the frame of the EC-JOULE Programme involving Danish and British research centers and meteorological services. They published the Danish activities and some preliminary results. Their model was based on the refinement of geostrophic wind estimates (outputs from high-resolution limited area weather prediction model (HIRLAM)) to a specific site, considering local effects as orography, roughness and obstacles (Machenhauer, 1988). Fellows and Hill investigated electrical load prediction (with an hourly time-step and a forecast horizon of up to 6 h) and wind turbine output power prediction (with a time-step of 10 min and a forecast horizon of up to 2 h) (Fellows and Hill, 1990). Watson et al. studied the reduction of fossil fuel costs, employing numerical weather prediction (NWP) and model output statistics (MOS). His model could predict wind speed and direction in an hourly time step and a forecast horizon of up to 18 h. They applied wind speed forecasts into a simulation of the UK grid system. They concluded that the NWP/MOS forecasts could significantly impact fossil fuel savings over persistence (Watson et al., 1992).
Tande and Landberg showed that neural networks could predict the output power from a single wind turbine using wind speed as a variable. They aimed to forecast the single value, ten steps ahead in a time series with 1 s data. They obtained only a minor improvement over the persistence model, pointing that further investigations should be carried out before any inferences about the model’s capability can be finalized (Tande and Landberg, 1993). Martin et al. proposed launching a Spanish project to develop an ad hoc operational tool for short-term forecast of local wind conditions and wind farms involving public and private research centers, a wind turbine manufacturer and electricity companies. The tool had a forecast horizon of up to 3 to 6 days and was based on local wind regime analysis, regional climatological analysis and wind farms parameterization (Martin et al., 1993). Jensen et al. proposed the wind power prediction tool (WPPT), developed by the Mathematical Modelling from the Technical University of Denmark (DTU), Department of Informatics and the ELSAM service area, a Danish utility. WPPT had a forecast horizon of up to 36 h and a half-hourly time-step. The tool was tested on seven wind farms. These were constructed based on an autoregressive model with output power as the primary variable and wind speed as the exogenous variable (Jensen et al., 1994).
Kariniotakis et al. investigated a fuzzy logic-based approach and neural networks. They found improvement over the persistence method in predicting the wind turbine’s output power with a prediction horizon of up to 2 h and a time-step of 10 min (Kariniotakis et al., 1996). Akylas et al.
tested different methods based on meteorological estimates and recorded data from wind masts to predict wind speed and translate it into wind turbine output power (Akylas et al., 1997). Nielsen
12
et al. proved that it is not reliable to use the persistence model as a reference model for a forecast horizon larger than a few hours. Instead, they proposed a weighting between the persistence and the mean of the power as an adequate reference model for all forecast lengths (Nielsen et al., 1998).
During the 1990s, wind forecasting methods evolved to consider a broader time horizon and incorporate the effects of terrain on wind speed. Forecasting of wind power from wind speed and wind turbine curve was defined. Modern approaches based on fuzzy logic and neural networks were developed. Moreover, the overall reliability of different forecast methods was studied during this period.
2.2.3 Since the 2000s
Sfetsos compared linear models with non-linear models (radial basis function network, feedforward neural networks, ANFIS models, Elman recurrent network, and neural logic network) to forecast mean hourly wind speed time series (Sfetsos, 2000). Watson et al. reported preliminary results of the online operation of the RISØ National Laboratory’s model. They informed that the initial results for 15 wind farms in Ireland were promising, but the model was not tuned for the specific sites (Watson et al., 2002). Giebel et al. proposed launching a 3-year project to integrate the models Prediktor and WPPT. The Danish Ministry of Energy funded the project. The final tool named Zephyr was supposed to get the advantages of both models. On the one hand, WPPT has better predictions for up to 6 h and the capacity to extend the meteorological service’s prediction horizon and handle changes in the input, for instance, changes in the weather model.
On the other hand, Prediktor had better predictions for horizons ranging from 6 h to the meteorological service’s forecast horizon and with the option of making forecasts even without available measurements (Giebel et al., 2002).
Martı´ et al. investigated the refinement of the outputs from the HIRLAM model to forecast a wind farm’s output power with reasonably complex terrain in Spain (Martí et al., 2003). Focken et al. studied the forecast of the aggregated output power of wind farms over several regions with a forecast horizon of up to 48 h. They detected a decrease in the prediction error due to spatial smoothing. They determined that the reduction in the error is much more sensitive to the size of the area of the wind turbines than to the number of sites when studying 30 wind farms in Germany (Focken et al., 2002). In December 2002, the International Energy Agency (IEA) held the Joint Action Symposium on Wind Forecasting Techniques. As the basic data set as input (wind turbine manufacturer power curves and meteorological forecasts), the output power forecast was made based on the wind speed forecasts and the manufacturer power curves. Considering an additional data set as input (the basic data set plus the measured power curves), better output power prediction was made, including wind direction forecasts. Further, considering the complete data set as input (the basic data set plus online measurements), the best output power forecast was possible through an ensemble of nine mathematical/statistical models (Sanchez et al., 2002).
Madsen et al. presented a procedure for standardizing the performance evaluation of short-term prediction models. They observed that using the persistence model as a reference model leads to over-optimistic conclusions about the model’s performance. They also reported the employment of the proposed procedure on the ANEMOS database and emphasized the need for enhancements on the methods for uncertainty assessment (Madsem et al., 2004). Bustamante et al. examined hourly wind speed prediction with two approaches, autoregressive models and neural networks
13
for time series. Two downscaling methods, i.e., dynamical and statistical, for very short-range time scales were studied. They concluded that statistical downscaling, which combines atmospheric fields simulated by some reanalysis projects, like ERA15 or ERA40, improves the results over a dynamic downscaling with a regional or mesoscale model to reach better resolutions for forecasts (Bustamante et al., 2004). Kariniotakis et al. presented limited results from the ANEMOS Project.
Based on six wind farms (one offshore, two in flat onshore terrain, two in complex terrain and one in highly complex terrain) distributed in the Northern and Western regions of Europe, they concluded that:
(i) the NWP spatial resolution was of high importance, especially in complex terrain;
(ii) the performance of the models greatly depended on the complexity of the terrain, showing reduced accuracy in complex terrain;
(iii) the prediction model errors were more pronounced in complex terrain;
(iv) the combination of different predictions should be considered to reduce the error (Kariniotakis et al., 2004).
Firat et al. (2010) proposed a statistical model based on the AR model for wind speed forecasting and independent component analysis. Using six years hourly wind speed of a wind farm in the Netherlands, they claimed that the proposed model could give higher accuracy than the direct forecasting methods for 2–14 h ahead (Hanifi et al., 2020). Bilal et al. designed an MLP network to predict the wind power of four wind farms in Senegal, West Africa. Wind speed was the primary input of their model, but they also assessed several input variables like temperature, wind direction, humidity and solar radiation. They concluded that air temperature has the highest impact on improving the accuracy of the wind forecast model (Bilal et al., 2018).
Wind forecasting methods continued to improve with the amount of data being analyzed. Factors such as temperature, humidity, solar radiation, wind direction, terrain roughness were studied extensively to improve the accuracy of wind forecasting methods. The wind energy forecasting models have gone through extensive improvement and expansion in recent years. Various types of commonly used intelligent predictors are organized into eight categories, i.e., ANN, ELM, SKM, FLM, AE, RBM, CNN, and RNN. Deep learning is a new method emerging in recent years that shows promising advances in wind energy forecast (Liu et al., 2019). As understood from these works, wind forecasting methods are undergoing constant changes and complexity. With new techniques and large amounts of data for the study, wind forecasting methods are getting better than ever.
2.3 Cases of overestimation of AEP predictions
In a DNV GL (2019) validation of pre-construction AEP predictions, the results are published for Great Britain, Ireland (including Northern Ireland), the UK offshore and South Africa. For Great Britain, the AEP predictions were overestimated by 3.1% based on data from 87 wind farms.
Considering a smaller amount of data, Ireland and South Africa had 4% to 5% overestimation, but the offshore wind farms in the UK were only 0.4% underestimated. According to DNV GL, the overestimations are most likely linked to shortcomings at the pre-construction energy assessment stage (DNV GL, 2019). The neglect of the wind-farm blockage effect is suggested to be a key reason for overestimating AEP predictions (Bleeg et al., 2018). Previously within the same project
14
as this study but based on a smaller amount of wind turbine data, production data from numerous existing wind farms in Sweden were compared with the pre-construction AEP predictions. The preliminary results were presented in (Möllerström and Lindholm, 2020), which showed a significant overestimation in the case of Sweden than for the British Isles and South Africa.
15
3 Research Question
In continuation to the work done by (Möllerström and Lindholm, 2020), utilizing the data available from Vindstat (“Wind Power Forecasting - Vindstat,” 2020), the production data of Swedish wind farms are compared with the AEP for a period from 1999 to 2016. Previously (Möllerström and Lindholm, 2020) have considered only wind turbines greater than 1.8 MW, whereas this study includes wind turbines from 55 kW to 3.3 MW. The different wind turbine models included in this study are listed in Appendix-I. Based on data available from wind turbines installed from 1988 onwards, this thesis aims to find out the effect of terrain type (open terrain, forest terrain and offshore), different turbine manufacturers, size of the wind farm has on the AEP of these wind turbines or farms. The AEP values are analyzed over time to see if there has been an improvement in the wind energy forecast as expected by improved wind forecasting methods over time. Further, as seen from the study by (DNV GL, 2019), this study examines if the overestimation of AEP is also a case for Swedish wind farms. This thesis report is made in parallel with a research paper, written together with the thesis supervisor Erik Möllerström (Möllerström et al., 2021) and which, in addition to the thesis, also incorporates a normalization to even out effects from wind turbine performance decline.
16
4 Methodology
A crucial step when developing a wind farm is making a reliable AEP prediction. The AEP prediction is made using wind-farm project planning software; for most cases, using a linear flow model, whereas CFD models are recommended for complex terrains. The chosen model evaluates wind data, which can be used to predict the AEP for a typical wind year for a planned wind farm if long-term correlated. A P50 value (50th percentile), which means a 50% probability of exceeding the predicted value for an actual wind year, can be acquired from these calculations. According to the standard (International Electrotechnical Commission, 2013), the predicted AEP (P50) is calculated as a mean value. The mean and median values coincide for a normal distribution of AEP over time; hence the use of mean value can be justified in this case. From the P50 value and expected uncertainties from different parameters, different P-values can be estimated. The wind resource evaluation and subsequent AEP prediction is an uncertain process. A proper assessment of the uncertainty is crucial for evaluating the risk with a planned wind energy development (Lackner et al., 2008).
4.1 Data and Software
The data used in this study is taken from the database Vindstat, which has been gathering production data, that includes downtime, from Swedish wind turbines since 1988 (Olauson et al., 2017). Vindstat data was used previously for evaluating the wind turbine performance decline in Sweden (Carta et al., 2013). Data from most Swedish wind turbines were included until 2016. The Swedish energy agency stopped its financial support in 2017, and Vindstat then introduced a member fee. Several large wind turbine owners stopped reporting after 2016, limiting the included wind turbines to nearly half for the subsequent periods. Besides production data, Vindstat includes the predicted AEP reported by the wind turbine owner or wind farms. This AEP (P50-value) corresponds to a typical wind year and is taken from the wind energy calculation from the planning phase (Möllerström and Lindholm, 2020).
For this study, all wind turbines that were still in operation in 1999 and reporting data to Vindstat were used. However, many might have been installed before 1999. In total, 2083 turbines were used for this study. The production data between 1999-2020 was then analyzed. The geographical spread of these turbines is shown in Figure 4-1. After excluding turbines according to the criteria described in “3.4 Data exclusion”, the number of wind turbines used in the analysis is given in Table 4-1. The different wind turbines categorized by model number, rotor diameter, rated power used in this analysis are shown in Appendix-I.
Table 4-1 Wind Turbine Data from Vindstat. Data excluded as per criteria in section 4.4
Description No data exclusion With data exclusion
Total wind turbines 2083 1727
Installed capacity (GW) 3.28 2.52
No of observations 235903 132683
17
Figure 4-1 Locations of wind turbines (green). Locations of index sites used for typical wind-year for normalization (red)
This study complements the work done by (Möllerström and Lindholm, 2020). A larger amount of wind turbines have been included in this analysis, and a more in-depth analysis is performed.
The results are divided into factors such as wind turbine manufacturers and wind farm sizes to see the impact of these factors. Time trends concerning the installation year have been analyzed to determine whether expected accuracy improvements in wind power forecast can be observed with time. The results were further divided by terrain type to determine whether the complexity of the
18
terrain affects the accuracy of AEP. Together with the previously published (Möllerström and Lindholm, 2020), this work is likely a unique case for Sweden.
4.1.1 ArcGIS
GIS is a tool that allows for spatial modeling in many fields and has recently been applied in the renewable energy sector. The geographic features of a wind farm are represented using spatial data such as topography, land cover, and wind resource. The wind turbines are represented by round dots, including information about the roughness of the surrounding terrain (Grassi et al., 2014).
World Imagery makes available one meter or better satellite imagery and aerial imagery in many parts of the world and lower resolution satellite imagery worldwide. The map includes 15m TerraColor imagery at small and mid-scales (~1:591M down to ~1:72k) and 2.5 m SPOT Imagery (~1:288k to ~1:72k) for the world. The map also features 0.5 m resolution imagery in the continental United States and parts of Western Europe from Maxar. Additional Maxar sub-meter imagery is featured in many parts of the world. In other parts of the world, imagery at different resolutions has been contributed by the GIS User Community. Very high-resolution imagery (down to 0.03 m) is available in select communities down to ~1:280 scale (“World Imagery - ArcGIS,” 2021). Figure 4-2 shows the interface of the GIS software with the different turbines in green dots and the index location in red dots on the map of Sweden, used in this study.
Figure 4-2 ArcGIS Mapping Software Interface
19
Figure 4-3 Wind Turbine Location, Index Location and Terrain as seen in ArcMap Software
As shown in Figure 4-3, each wind turbine’s terrain was identified and categorized as open, forrestal and offshore terrains. Further, each wind turbine was allocated with the nearest wind index location. As shown in the Figure 4-3, for wind turbine number 1602, the index location is 1, and the terrain is open terrain. The distance from each wind turbine to the index location is also identified, and in the above case, it is 5 km. This is done for all the wind turbines in the study.
Further, based on the operation start date, name of the wind farm, manufacturer and model of the turbine, and the distance to nearby turbines, the number of wind turbines in a farm was identified.
4.2 Analysis Method
The production data from Vindstat is normalized to correspond to the P50 prediction and the wind energy available for a typical year as an initial step. Then, the monthly production data were re-calculated to the wind index corrected annual production (WCP) and then compared to the P50 values. MATLAB was utilized for all the below calculations.
4.2.1 Normalizing of production data
The P50 value is assumed as the availability that the owner can count on (the manufacturer guaranteed availability, typically 97% (Olauson et al., 2017)). The time availability is the percentage of a given period that a wind turbine is available for operation. For a given month, it was calculated using the following formulae (Möllerström and Lindholm, 2020).
Index Location-1
Wind Turbine Littra 1602
20
𝑇𝑇𝑇𝑇𝑇𝑇 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎, 𝐴𝐴𝑎𝑎 =𝑇𝑇𝑇𝑇−𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 Eqn. 4-1 Tm is the total number of hours in a particular month.
Td is downtime for hours in the same month (Möllerström and Lindholm, 2020).
The availability of the same month was used to calculate the electricity yield for the typical industry standard that is 97% (Astd) using the following formulae (Möllerström and Lindholm, 2020).
𝐸𝐸𝑎𝑎𝑇𝑇𝐸𝐸𝑎𝑎𝐸𝐸𝑎𝑎𝐸𝐸𝑎𝑎𝑎𝑎𝑎𝑎 𝑎𝑎𝑎𝑎𝑇𝑇𝑎𝑎𝑦𝑦, 𝐸𝐸𝑎𝑎 = 𝐸𝐸∙𝐴𝐴𝐴𝐴𝐴𝐴𝑇𝑇∙𝜂𝜂𝐴𝐴
𝐴𝐴𝐴𝐴 Eqn. 4-2
E is the electricity yield for a given month.
ηt is the transformer efficiency which is assumed to be 99%
P50 value is measured at the grid connection 4.3 Wind Index Correction
The short-term wind measurement data considered when calculating the predicted AEP of a proposed wind turbine should be correlated using a long-term wind dataset due to the interannual variations in wind speed (Carta et al., 2013). For a long time series, as in this case, the insecurity of the correlation may exceed that of the validation between the data and a typical wind year (Zhang et al., 2019). However, when analyzing a vast number of wind turbines, they must have data for the same long period. Since different periods for different wind turbines are considered in this study, the production data must be individually correlated before analyzing against the AEP.
The normalization of the data was done by extracting the monthly correlation indices for 77 sites spread across Sweden. Figure 4-1 shows the location of these indices in red round dots. These correlation indices were calculated using the ERA5T reanalysis, which performed well for the often complex terrain sites of Swedish wind turbines (Olauson, 2018, p. 5). These correspond to the wind energy at 100 m height for each site for the actual month, compared to that same month for a typical wind year. The yearly mean of all 77 index locations is shown in Figure 4-4. For one wind turbine, all production data in a month is normalized for availability (Ea) and plotted against the correlation index of the same month with the index location closest to the specific wind turbine (In). This is shown in Figure 4-5 for one of the wind turbines. The wind-index-normalized production per month (En) was achieved from a linear regression line for In = 1. This is done for
21
all the wind turbines. The wind-index-corrected annual production (WCP) was then calculated using the following equation (Möllerström and Lindholm, 2020).
𝑊𝑊𝑊𝑊𝑊𝑊 = 12 ∙ 𝐸𝐸𝐸𝐸 Eqn. 4-3
The error between the production WCP and the predicted P50 (AEP) from the project developing phase can then calculated using the following equation (Möllerström and Lindholm, 2020).
𝑇𝑇𝐸𝐸𝐸𝐸𝑒𝑒𝐸𝐸 = 𝑃𝑃50−𝑊𝑊𝑊𝑊𝑃𝑃
𝑊𝑊𝑊𝑊𝑃𝑃 Eqn. 4-4
A positive error indicates that the P50 value was an overestimation of the actual AEP and a negative error indicates an underestimation.
Figure 4-4 Yearly mean correlation index of all 77 sites
22
Figure 4-5 All monthly production data of a single wind turbine normalized for availability (Ea) is plotted against the correlation index for the same month with the index location closest to the same wind turbine (In)
4.4 Data Exclusion
Wind turbines with rated power less than 50 kW were removed from the analysis. Wind turbines that were not in operation from 1999 were removed from the analysis. From the production data, the months in which the sum of the downtime of the generator and the reported operation time were greater than the hours of that month were excluded to avoid errors. Months with availability below 90% were also removed. To remove the influence of corrupted data in the analysis, the points which are significanlty away in the production-index plotline must be removed. This was done using a WCP filter of +/-15%. The index-normalized monthly production is compared to the WCP and excluded according to the following equation (Möllerström and Lindholm, 2020).
0.85 ∙ 𝑊𝑊𝑊𝑊𝑊𝑊 > 12 ∙𝐸𝐸𝐸𝐸𝐼𝐼𝐼𝐼 > 1.15 ∙ 𝑊𝑊𝑊𝑊𝑊𝑊 Eqn. 4-5 The WCP was re-calculated without the filtered values. If the remaining month-production values were fewer than 12 or their coefficient of determination (r2) was less than 0.85, the corresponding wind turbine was removed from further analysis.
23
5 Results
This study has analyzed a total of 1727 wind turbines with a total installed capacity of 2.52 GW.
The production data utilized in this study is for wind turbines that were in operation since 1999, and wind data until 2016 has been analyzed. Since 2017, only a few wind turbine owners have reported data to Vindstat; these data were excluded from the analysis. For each wind turbine, the monthly wind power production data was normalized and re-calculated according to Eqs. (4.1-4.3) and comparing it to the given P50 in Eq. (4.4), the errors of the pre-construction AEP predictions compared against production-based WCP were calculated. The mean error based on all wind turbines in the study was 11.9%, which means that the predicted AEP from the planning phases of the studied wind turbines was overestimated. The mean overestimation is more significant than the 7.1% previously shown for the same dataset but only looking at wind turbines with a rated power of 1.8 MW and more and installed after 2005 (Möllerström and Lindholm, 2020). The results also emphasize the more significant error for Swedish wind power projects than that previously understood for the British Isles and South Africa (DNV GL, 2019).
Figure 5-1 Error of P50-evaluations compared to the production-based WCP depending on the installation year plotted with the standard deviation (blue) and the number of wind turbines (red).
In Figure 5-1, the error is divided by each installation year of the wind turbines (total 1727 WTGs).
The AEP predictions have, with the exceptions of 2013 and 2016, been overestimated. Although the accuracy can be said to have improved from 1990 until around 2005, as can be seen by the
24
constant decrease in mean error and the standard deviation, it has stagnated later. Early improvement is anticipated due to the relatively small-scale wind power projects in the early 1990s and immature forecasting methods resulting in higher error. Then the consecutive implementation of wind-farm project planning software and improvement in forecasting methods influenced accuracy improvement in AEP prediction. This can be attributed to the development in computing systems during this period. The accuracy was highest for wind turbines installed from 2011 to 2013, followed by a drop in the accuracy level in 2014 and 2015, which is interesting to note. The lack of improved accuracy in recent years in forecasting wind power contradicts improvements in the field of forecasting such as access to better long-term corrected wind data, improved flow models, improved tools for sector-based zero-plane displacement, more detailed terrain data, refined measurement methods, improved standards, and a better understanding of climatic effects.
Table 5-1 shows the error based on all wind turbines and divided on terrain type, wind turbine manufacturer, and the number of wind turbines in the wind farm. The error is more significant for the less-complex flat terrain than for forestall terrain, indicating that the complexity of the forest terrain is not the main reason for the error. The accuracy was higher for offshore turbines, but this is based on a relatively small amount of data. Dividing the error between the most common wind turbine manufacturers in the analyzed data, the error was slightly larger for Enercon turbines than for Vestas turbines. However, both cases were more accurate than the turbines of other brands combined. Regarding wind farm size, the accuracy was higher with more turbines. In Table 5-1, the error divided on installation decade can also be seen, and although still overestimated, the error was comparatively less for the 2020s with 6.3%.
Table 5-1 Results
Classification Error (%) Turbines Power
(MW) Observations
All data 11.9 1727 2527 132683
Terrain
Flat 12.8 955 1063 84326
Forestall 11.7 680 1268 41626
Offshore 4.9 92 196 6731
Manuf.
Vestas 10.1 837 1226 67762
Enercon 10.7 440 650 37500
Other 16.5 450 651 27421
25
Classification Error (%) Turbines Power
(MW) Observations
Farm size
Single 14.0 383 300 34144
2-4 12.3 578 726 49152
5-10 11.3 389 649 29614
>10 10.0 377 854 19773
Install.
Decade
1990s 20.2 436 199 43818
2000s 12.0 590 793 52890
2010s 6.3 691 1533 35697
In Figure 5-2, the development of the AEP prediction accuracy over time can be seen divided on terrain type, manufacturer and wind farm size. Figure 5-3, Figure 5-4 and Figure 5-5 show the same plots with standard deviation and the number of wind turbines installed each year. The wind turbine data in all these plots are plotted for the installed year as it is interesting to note the development over the years of newer installed wind turbines.
26
Figure 5-2 The time development of the AEP prediction accuracy divided on terrain type, manufacturer and wind farm size.
27
Figure 5-3 The development of the AEP prediction accuracy with time (blue) divided on terrain type with the number of wind turbines for each year (red).
As seen in Figure 5-3, results divided on terrain, installations before 2004 had a higher accuracy for flat terrain than forestall terrain, while installations in forestall terrain have been slightly more accurate after 2004. The improvement over time for installations in forestall terrain is probably an effect of improved models and data used for the AEP prediction. Even though the forrestal terrain has a turbulent wind field, from the results, it can be understood that the prediction of AEP for forrestal terrain has improved despite its complexity. In Sweden, the number of wind turbines in the forrestal area has increased over the years, especially since 2008. Despite this increase in the number of wind turbines, after the year 2004, the error remains consistent.
The offshore turbines showed an overestimation in the year 1997. However, the lack of good data for offshore wind turbines has inhibited coming to a relevant conclusion. Wind turbines in offshore locations were prominently installed in 2007 and later in 2013. More installation of offshore wind farms are required to provide enough data for future analysis.
28
Figure 5-4 The development of the AEP prediction accuracy with time (blue) divided among the manufacturers with the number of wind turbines installed for each year (red).
Vestas has a vast history in wind energy compared to Enercon, which entered the Swedish market in 1995. Figure 5-4 shows the influence on the manufacturer for the AEP predictions; projects with Vestas turbines had large overestimations around the year 2000 but have otherwise had a higher accuracy than for the case of Enercon. Enercon showed considerable overestimation during 1997 and 2003 and slight overestimation in 2009, 2010 and 2014.
However, for other years it showed a consistent estimation. For the other wind turbines, we can note that initial years had constant improvement due to the improvements in forecast methods but are nearly constant for the later years of turbine installation. It is interesting to note that in 2012 and 2013, there was underestimation in AEP. It can also be noted here that the wind industry in Sweden has developed significantly with the most number of wind turbines installed from the manufacturers, Vestas and Enercon, since 2007. Table 5-1 shows that the installed capacity of wind power has significantly increased over the years in Sweden based on the data analyzed with
29
199 MW in the 1990s, 793 MW in the 2000s and 1533MW in the 2010s. This can also be attributed to the fact that most manufacturers have developed higher-capacity wind turbines in recent years.
Figure 5-5 The development of the AEP prediction accuracy with time (blue) divided on wind farm size with the number of wind turbines installed for each year (red).
Dividing the results on the size of wind farms as seen in Figure 5-5, there is no apparent difference in accuracy between farm sizes, with wind farms of more than ten turbines having the highest overall accuracy. If wind-farm wake effect was the main contributor to the overestimations of AEP predictions, a higher overestimation could have been anticipated in wind farms than for single turbines. However, that is not the case here. It is reasonable to assume here that as wind forecasting models improved over the years, the case of wake effect has been taken into relevant consideration
30
in AEP. Moreover, it can be understood from Figure 5-6 that the mean error is decreasing with the increase of the size of the wind farm, which is interesting to note.
Figure 5-6 Mean error for wind farms of different size
