You are on page 1of 17

Journal of Cleaner Production 313 (2021) 127933

Contents lists available at ScienceDirect

Journal of Cleaner Production


journal homepage: www.elsevier.com/locate/jclepro

Assessment of wind energy potential using reanalysis data: A comparison


with mast measurements
Rajat Kanti Samal
Department of Electrical Engineering, Veer Surendra Sai University of Technology Burla, Odisha, 768018, India

A R T I C L E I N F O A B S T R A C T

Handling editor: Panos Seferlis Reanalysis datasets are widely used for wind resource assessment (WRA) when measured data is unavailable.
Due to variation in wind resource, it is essential to establish the suitability of such datasets for the selected region
Keywords: by comparison with available measured data. This work is a pioneering attempt at performing detailed com­
MERRA-2 parison between NASA MERRA-2 and 50 m mast measured wind speed data in Odisha state of India. Initially, the
Measured wind speed
descriptive statistical measures are computed followed by computation of relevant error measures and corre­
Correlation
lation analysis between the datasets. In the next step, selected marginal distributions are fitted and goodness-of-
Marginal distributions
Wind power density fit statistics are compared. Finally, the wind power density computed using data from both the sources are
examined. It is found that the datasets can be considered to be in agreement only for longer durations such as a
year; large differences are observed in hourly, monthly and seasonal variations. Wind power density computed
from the two sources widely vary across months which is expected to affect the utility of MERRA-2 data for
computing monthly and seasonal energy capture. The work is novel both in the details of the analysis using
measured wind resource and the application region, that is, the east coast of India. The results of the analysis are
expected to enrich the existing literature on wind resource probabilistic modelling and utility of MERRA-2 for
WRA studies.

concept of wind power forecasting (WPF) is only relevant for day to day
operation of the wind power plant. Primarily, WRA consists of proba­
1. Introduction bilistic assessment of the wind speed data collected over a long time span
of several years. Ideally, this multiple year data must be measured at a
Renewable energy deployment is one of the significant enablers for suitable hub height of at least 50 m. However, collection of site
achieving United Nations Sustainable Development Goals. Wind energy measured data over a multi-year period is cost-intensive and commercial
is amongst the most preferred non-conventional energy options in the developers generally collect data for only a couple of years. Even over
world energy portfolio and is expected to significantly contribute to such short periods, the measured data generally have a lot of measure­
emission reduction (Hernández et al., 2019). The expected energy cap­ ment gaps. Another alternative to the use of historical datasets is the
ture and hence the economic and financial viability of a wind power data generated from Numerical Weather Prediction (NWP) models
project is invariably dependent on the wind regime of the selected which are effectively 2 km–5 km resolution physical models. However,
location, especially the wind speed. Specifically, combination of wind these are computationally intensive. Therefore, decisions based on one
and solar generation are projected to be the energy solutions in the near year of measured data are not uncommon and in most of these cir­
future. A number of industrial organizations are investing on wind en­ cumstances and the measured data is supplemented by satellite based
ergy to meet their electricity needs (Zare Oskouei et al., 2021). The issue reanalysis data. A common assumption underlying the use of reanalysis
becomes more complex as a result of climate change which is expected datasets is that they have the same characteristics features as the
to alter the wind speed probability distributions. Therefore, any new real-time data. Nevertheless, evaluation of consistency between rean­
wind power project development is preceded by detailed wind resource alysis and measured datasets, at least over a time span of one year, can
assessment (WRA) for evaluating the suitability of the location (Chandel add more authenticity to wind resource assessment performed using the
et al., 2014a). In this direction, attempts have been made to characterize former.
wind resource variations at global and continental levels to aid in wind
power potential estimation (Campisi-Pinto et al., 2020). The related

E-mail address: rajatksamal82@gmail.com.

https://doi.org/10.1016/j.jclepro.2021.127933
Received 1 February 2021; Received in revised form 6 May 2021; Accepted 12 June 2021
Available online 16 June 2021
0959-6526/© 2021 Elsevier Ltd. All rights reserved.
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Abbreviations PD Probability Distribution


PDF Probability Density Function
B–S Birnbaum-Sanuders distribution WMS Wind Monitoring Station
CDF Cumulative Distribution Function WRA Wind Resource Assessment
DA Daily Average WPD Wind Power Density
HA Hourly Average
GEV Generalized Extreme Value distribution Statistical Measures
IG Inverse Gaussian distribution CV Coefficient of variation
LL LogLogistic distribution rMAE relative mean absolute error
LN Lognormal distribution rRMSE relative root mean square error
MES50m Measured wind speed data at 50 m measurement mast GOF goodness-of-fit tests
MES20m Measured wind speed data at 20 m measurement mast chi2 χ 2 Statistic
MER50m MERRA-2 50 m wind speed K–S Kolmogorov-Smirnov statistic
MER10m MERRA-2 10 m wind speed

1.1. Data sources Online availability of this data is ensured by Goddard Earth Sciences
(GES) Data and Information Services Center (DISC). The data files are
Reanalysis datasets are generally based on global circulation models commonly arranged in scientific data format (netCDF-4). There are
and measured data at specific locations. A number of such datasets are provisions to download the data for a specific location also and detailed
available which are mostly from country or continent level meteoro­ procedures are available for the same. It is worthwhile to mention that
logical organizations. Prominent among the reanalysis datasets are the MERRA-2 has a number of data collections which includes instantaneous
NCEP dataset by National Center of Environmental Prediction (NCEP), and time averaged data.
ERA dataset by European Center for Medium Range Weather Forecasts
(ECMWF) and JRA by the Japanese Meteorological Agency (Carvalho 1.1.2. Measured wind resource data
et al., 2014). In (Chancham et al., 2017), the NCEP/NCAR (National Wind resource data measured at weather station masts are the most
Center for Atmospheric Research) database and measured data are used reliable. Light Detection and Ranging (LiDAR) can be used for measuring
for development of wind resource maps over a wide area convering the wind speed beyond expected hub heights, though the cost of deployment
Gulf of Thailand. WRA in three locations of Republic of Djibouti is is very high limiting such applications to industrial wind power devel­
performed in (Dabar et al., 2019) using NCEP-CFSR and ERA5 reanalysis opment projects (Sharma et al., 2019). Measured data using both LiDAR
data in conjunction with Weibull distribution. However, detailed com­ systems and anemometers are utilized in (Li and Yu, 2018) to compute
parison between reanalysis and measured data sources are not per­ wind power potential in offshore and onshore regions near Lake Erie
formed in the above mentioned works. NASA’s MERRA (Modern Era where statistical analysis includes determination of Weibull shape and
Retrospective Analysis for Research and Applications) is the most widely scale parameters. ArcGIS software along with meteorological data is
used reanalysis dataset. In fact, it is shown in (Kubik et al., 2013) that utilized in (Bina et al., 2018) for economic analysis of future wind power
simulation obtained from MERRA is comparable and sometimes better plants in northwestern Iran where Weibull distribution was fitted to
than the simulations derived from measured data, however, the com­ actual meteorological data. The impact of different geographical cli­
parisons are based on only correlation and error analysis. MERRA data matic conditions on wind power potential is analyzed in (Li et al., 2018)
for the period of 1998–2017 is compared with measured data from using measured data from six locations in China with the aid of Weibull
several wind farms and meteorological stations to estimate country wide distribution. Wind Atlas Analysis and Application Program software
wind profile in (Ren et al., 2019). Although the above work does include along with meteorological data is utilized in (Chandel et al., 2014b) for
validation of MEERA-2 using measured data, comparisons across economic analysis of future micro-scale wind power plants in Himalayan
months and seasons are not provided. MERRA datasets are also found region of India. However, mast measured data are available only for a
useful for detailed technical investigations such as turbine ageing few specific locations and are generally proprietary. Wind power project
(Staffell and Green, 2014) and regional wind power production (Olau­ developers make a lot of investment and assign a lot of resources for
son and Bergkvist, 2015), although the conclusions could have been collecting such data and hence do not provide them for free. Sometimes,
more robust with inclusion of detailed comparisons between reanalysis certain organizations such as the National Institute of Wind Energy
and measured values. In fact, if such validations are available for large (NIWE), Govt. of India provide mast measured data at a nominal cost. A
number of locations across the globe, wind speed data obtained from a comparison between MERRA-2 and the measured wind data can provide
combination of global climate models can be utilized to study the vari­ useful insights regarding the utility of the former so that multi-year
ation in availability of wind resource in the past and future due to global reanalysis data, after suitable modifications, can be utilized for WRA
climate change. studies and subsequent wind power project development.

1.1.1. MERRA 1.2. Uncertainty modelling of wind resource


NASA MERRA data are available from Goddard Earth Observing
System Model (GEOS) with Atmospheric Data Assimilation System Selection or assumption of a wind speed distribution is one of the
(ADAS). Specifically, MERRA refers to the first version of NASA atmo­ prerequisites for pursuing with WRA studies. The steps involved essen­
spheric reanalysis datasets which started from the year 1980 (Rienecker tially consist of modelling of wind speed characteristics of the concerned
et al., 2011). This first version was discontinued in 2008 and was region by means of probability distributions. The focus is on multi-year
replaced by the second and improved version known as MERRA-2 macro-analysis of the historical wind speed for the specified location. A
(Modern-Era Retrospective, 2018). The new version incorporated the number of past research works have focussed on probabilistic modelling
advances in newer satellite instruments and novel data collection of wind resource and a common consensus that has emerged from these
methodology. The MERRA-2 dataset is available on a grid with 576 works is that no single probability distribution can be considered to
longitudinal 361 latitudinal points with a resolution of 0.6250 × 0.50. represent the wind speed uncertainty. A detailed review of some of the

2
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Table 1
Wind speed probability distributions.
PD PDF f(v) CDF F(v) P1 P2 P3
( (v)k ) ( ( ) )
Weibull v k− 1 v k k c –
k exp − 1 − exp −
ck c c
Gamma 1 ∫ va− 1 a b –
va− 1 exp( − v /b) exp( − v /b)dv
b Γ(a)
a
{ } ba Γ(a)
( )
LN 1 − (ln(v) − μ)2 ln(v) − μ μ σ –
√̅̅̅̅̅̅ exp Φ
vσ ̅̅̅̅̅̅̅̅̅̅
2π 2σ2 σ
√ (√̅̅̅ ( ))
IG λ { λ } λ v μ λ –
exp − (v − μ)2 Φ − 1 +
2 π v3 2μ2 v ( v )μ ( √̅̅̅ (
2λ λ v ))
exp Φ − +1
[ ] μ v[ μ ]
TN 1 (v − μ)2 ∫v 1 (v − μ)2 μ σ –
√̅̅̅̅̅̅ exp − for v ≥ 0 √̅̅̅̅̅̅ exp − dv
I(μ, σ)σ 2π 2σ[2
]
0
I(μ, σ)σ 2π 2σ 2
∫∞
1 (v − μ)2
I(μ, σ) = √̅̅̅̅̅̅ exp − dv
[σ( )2π 0 ( ) ] 2σ[
2
BS 1 β 1/2 β 3/2 1 (v β )] [ {( )1/2
1 v (β)1/2 } ] β γ –
√̅̅̅̅̅̅ + exp − + − 2 Φ −
2 2πγβ ( v ) v 2γ β v
2 γ β v
Logistic v− μ 1 μ σ –
exp (v − μ)
σ
{ (v − μ) }2 1 + exp
σ 1 + exp σ
( σ )
LL ln(v) − μ 1 μ σ –
exp ( )
σ ln(v) − μ
{ ( ) }2 1 + exp
ln(v) − μ σ
σ 1 + exp
σ ( ( )
Nakagami 1 ( μ )μ 2μ− 1 μ ) γ μ, v2
μ μ ω –
v exp − v2 ω
Γ(μ) ω ω
⎡ ⎤ ⎡μ)
Γ( ⎤
GEV 1 1 1 k σ μ
1 ⎢
( v − μ)− k ⎥( v − μ)− 1− k ⎢
( v − μ)− k ⎥
exp⎣ − 1 + k ⎦ 1+k exp⎣ − 1+k ⎦
σ σ σ σ

Thus it can be inferred that there is an implicit assumption in the above


Table 2
mentioned works about the universal applicability of Weibull distribu­
Descriptive statistics for year and seasons for MES50m.
tion when wind speed data is utilized.
Period Mean (m/s) CV Skewness Kurtosis Nevertheless, detailed investigations on the suitability and signifi­
Year 5.85 0.59 3.17 43.91 cance of other PDs such as Gamma, Lognormal, Inverse Gaussian (IG),
Summer 7.08 0.42 − 0.07 2.31 truncated Normal (TN), Logistic, LogLogistic (LL), Nakagami and
Rainy 6.03 0.48 0.43 2.88 Birnbaum-Saunders (BS) distributions have been performed (Moham­
Autumn 5.87 0.79 4.65 49.71
Winter 4.37 0.46 1.08 6.55
madi et al., 2017). The above mentioned PDs are categorized as
one-component marginal distributions and have been shown to be quite
suitable for a majority of locations. The other alternative is the use of
two-component mixture distributions. Examples are weighted combi­
Table 3
nations of Weibull-Weibull and Gamma-Weibull, TN-Weibull and TN-TN
Descriptive statistics for year and seasons for MER50m.
mixture distributions (Samal and Tripathy, 2018). However, probably
Period Mean (m/s) CV Skewness Kurtosis due to their complex mathematical representation, mixture PDs, have
Year 5.64 0.52 0.88 5.04 not been utilized for power system studies. Further, though wind di­
Summer 6.96 0.38 − 0.09 2.48 rection is not an important parameter in wind energy studies, joint PDs
Rainy 6.72 0.42 0.21 2.67
involving wind speed and direction have been the subject of some
Autumn 5.00 0.61 2.36 12.54
Winter 3.82 0.47 0.27 2.50 research works (Cook, 2019). An alternative method of WRA studies is
the use of computational fluid dynamics (CFD) models (Valsaraj et al.,
2020) which also requires application of probability distributions. CFD
recent significant works on wind speed probabilistic modelling is pro­ models can be utilized to extrapolate the measured data across various
vided below which eventually establishes the context of this work. The complex terrains using coupled modelling frameworks (Duran et al.,
most popular and widely used probability distribution (PD) is the Wei­ 2020). A major shortcoming of the above mentioned works is that the
bull distribution (Wang et al., 2018; Murthy and Rahi, 2017). It is said distributions have not been utilized for establishing validity of
worthwhile to note that power system studies involving the investiga­ reanalysis data.
tion of the impact of wind power on power system operation generally
use Weibull distribution for wind power scenario generation (Samal and 1.3. Summary of research gaps and contributions
Tripathy, 2019). Apart from above mentioned simulation studies, Wei­
bull distribution also forms a significant portion of works utilizing data The following research gaps are observed from the detailed literature
from measurement masts in addition to remote sensing data. A review of review.
the Weibull parameter based wind resource maps on coastal Africa is
performed in (Olaofe, 2018) with the help of cross-calibrated multi-­ 1. Most of the research works utilizing reanalysis data such as MERRA-2
platform surface wind based satellite observations. In (Sharma et al., are conspicuous in the absence of establishing the validity of such
2019), Lidar measurements are utilized to extrapolate the wind shear datasets by comparing with measured data at least for one year.
index for offshore WRA studies with the aid of Weibull distribution. Further, when such comparison is performed, it is done at macro
Continental scale assessments such as those detailed in (Elsner, 2019) level without detailed comparisons at seasonal and monthly
involve computation of power potential and capacity factors and are intervals.
quite dependent on the use of Weibull distribution and WAsP software.

3
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 1. Comparison of mean and coefficient of variation.

Fig. 2. Comparison of skewness and kurtosis.

2. In past research works which do include comparison between shared responsibility. In fact, this has led to the incorporation of
MERRA-2 and measured wind data, the evaluation metrics are Intended Nationally Determined Contributions (INDC) in the Paris
mostly correlation and error measures such as coefficient of deter­ agreement. In order to have maximal impact, all regions of the world
mination and root mean square error. Probability distributions, should contribute towards development of non-conventional and
which form the core of wind resource stochastic modelling do not clean sources of energy. From the detailed literature review, the
form the part of the comparison in any of the works reviewed by the author could not find any scholarly published work comprehensively
author. evaluating the validity of MERRA-2 for WRA studies in the Odisha
3. Finally, as acknowledged by the United Nations Conference of coast of India. Due to the abundant wind resource in this region and
Parties from time to time, climate change mitigation is a common but the current focus of the government to develop new wind power

4
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Table 4 potential studies. Seasonal and monthly variations of wind resource can
Error Measures for year and seasons. provide necessary inputs for efficient wind power project planning and
Period Hourly Average Daily Average implementation. Further, wind resource modelling in the selected
location can be extrapolated to compute the wind power potential for a
rMAE (%) rRMSE (%) rMAE (%) rRMSE (%)
large swathe of east coastal belt in the Odisha state of India and can also
Year 33.57 49.22 21.44 32.26 form a basis for offshore wind resource assessment.
Winter 42.44 55.52 24.14 33.20
Summer 24.04 30.91 11.19 15.20
Rainy 31.39 40.28 21.11 27.15 2. Description of the location and data
Autumn 40.91 69.36 32.28 49.09

Wind speed distributions are crucial for estimating the energy yield
projects, this region is expected to contribute significantly to wind from a wind farm. Estimation of parameters of a probability distribution
power development and consequent reduction of global carbon not only depends on the parameter estimation methods, but also on the
footprint thus requiring detailed investigation on suitability of quality of the data (Jung and Schindler, 2018). Availability of daily
MERRA-2. average (DA) data for longer time span is comparatively easy and have
been utilized for evaluation of suitability of distributions. However,
In view of the above mentioned research gaps, the specific contri­ results obtained using DA data may not faithfully represent wind speed
butions of this work are as follows. uncertainty and its use is mostly due to unavailability of multi-year data
with higher temporal resolutions. Hourly average (HA) data is the most
1. A comparison is performed between the 50 m mast measured data useful for WRA studies since most wind speed PDs have been found to
and MERRA-2 data to analyse the similarity in descriptive statistical model HA data. In majority of the cases, availability of hourly average
parameters. wind data implicitly implies the availability of 10-min averaged data.
2. Hourly average data from both the sources are compared and error The major issue confronting researchers is that 10-min average
measures and correlation are computed to verify the closeness of the measured data sources are scarce due to the associated cost. Even when
sources. such data sources are available, for example, from national level
3. Ten selected marginal distributions are fitted to both the measured
and MERRA-2 wind speed and their suitability to model wind speed Table 5
uncertainty is established. Correlation for year and seasons.
4. Wind power density obtained from both the sources are compared to Period MES20m-MER10m MES50m-MER50m
find the suitability of MERRA-2 for WRA studies. HA DA HA DA

Year 0.6920 0.8586 0.6017 0.7651


The contributions of this work are expected to provide significant Winter 0.1160 0.5070 0.2276 0.5136
insights into the nature of wind resource to researchers across the globe. Summer 0.7010 0.8880 0.7089 0.9001
The new findings on the suitability of MERRA-2 and wind speed prob­ Rainy 0.6143 0.7634 0.6678 0.8115
ability distributions can be used by researchers performing wind power Autumn 0.7500 0.8846 0.5361 0.6918

Fig. 3. Error measures across months.

5
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 4. Correlation across months.

Fig. 5. GOF Statistics for yearly wind speed.

meteorological service providers, the quality of the data is very poor. of coastal plains having coast line of 450 km. The state comes under the
The availability of commercial grade wind resource data is normally for eastern region of the Indian Power Grid and has a total installed capacity
a few years and the wind project developers work with one year of of 8567 MW. This energy mix consists of 5894 MW of coal fired thermal
measured data most of time. Commercially measured data available at a power and 2151 MW of hydro power. The share of renewable energy
cost is far superior in quality as compared to those provided by meteo­ (solar) is meagre at only 521 MW. Although the state is expected to have
rological organizations even though the time span is less. about 3000 MW of wind power potential (Odisha Renewable Energy D;
Odisha is one of the major states in the east coastal belt of India National Institute of Wina), no commercial wind power project is either
having a 42 million population and a share of 4.87% of the total area of operating or under development in the state. The reason for this slow
India. The state borders Bay of Bengal in the eastern part which consists growth of renewable energy development is because of energy surplus

6
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Table 7
Distribution Parameters for Seasonal Wind speed (MES50m).
Season PD P1 P2 P3 pChi pKS

Summer GEV − 0.25 2.97 5.96 0.0000 0.0001


Rainy Weibull 2.17 6.79 – 0.0435 0.0846
Nakagami 1.11 44.74 – 0.0228 0.0144
GEV − 0.15 2.62 4.85 0.4589 0.6131
Autumn LN 1.54 0.71 – 0.0368 0.0000
LL 1.56 0.37 – 0.0475 0.0310
Winter Logistic 4.27 1.05 0.0000 0.2822

Table 8
Distribution Parameters for Seasonal Wind speed (MER50m).
Season PD P1 P2 P3 pChi pKS

Summer Normal 6.96 2.68 – 0.0000 0.0323


GEV − 0.30 2.70 6.04 0.0032 0.1102
Rainy Weibull 2.55 7.56 – 0.3165 0.4132
Normal 6.72 2.82 – 0.0002 0.2012
GEV − 0.21 2.69 5.63 0.2819 0.7556
Autumn Gamma 3.08 1.62 – 0.0000 0.0004
Fig. 6. Fitted distributions for yearly measured wind speed. GEV 0.07 2.05 3.66 0.0000 0.0099
Winter Weibull 2.24 4.31 – 0.0008 0.0668
Nakagami 1.15 17.84 – 0.0000 0.0146
GEV − 0.19 1.69 3.11 0.0012 0.1176

in Odisha, highest mean wind speed of 5.72 m/s pertains to Pradeep


WMS and hence data from this WMS is chosen for this work. The exact
location of the wind monitoring station is 20015′ 01.4′′ East and
86038′ 53.1′′ North and the mast height is 80 m. The data is commercially
made available for purchase by wind project developers and also for
academic institutions at a discounted cost. The data used in the current
work is purchased from NIWE under institutional use terms and is of one
year time span, that is, March 2013 to February 2014. The NIWE
measured data used in this work are measured at 20 m and 50 m mast
heights. From MERRA-2, the vert M2T1NXSLVvert data product is uti­
lized which is basically one hourly averaged data for 2 m, 10 m and 50 m
heights. The 50 m MERRA-2 (MER50m) and WMS measured data
(MES50m) are primarily compared. The 10 m MERRA-2 and 20 m
measured data are also used and are referred to as MER10m and
MES20m respectively. The comparison is primarily for hourly average
(HA) wind speed with occasional references to daily average (DA)
values.

Fig. 7. Fitted distributions for yearly MERRA-2 wind speed.


3. Methodology

The relevant background of the methods utilized in this work are


Table 6
described in the following subsections.
Distribution Parameters for Yearly Wind speed.
PD MES50m MER50m
3.1. Measures of comparison

P1 P2 P3 P1 P2 P3
The preliminary measures of comparison of two sets of data sources
Weibull 1.80 6.57 – 2.01 6.36 – are the measures of descriptive statistics, namely the mean and coeffi­
Gamma 3.02 1.94 – 3.26 1.73 – cient of variation (CV). Additionally, skewness and kurtosis of the data
Nakagami 0.91 45.90 – 1.02 40.37 –
GEV 0.03 2.48 4.35 − 0.05 2.44 4.33
can indicate general measures of shape of the observations over a period.
However, the above measures are only used for macro-level comparison
and generally do not provide much information on the discrepancy or
status of Odisha. However, the detrimental environmental impacts of similarity of the data. Detailed information can be obtained by the use of
thermal and hydro power have not gone unnoticed by the policy makers mean square error (MSE), root mean square error (RMSE) and mean
and currently there is shift in policy towards renewable energy, mostly absolute error (MAE) performance metrics (Adedeji et al., 2020). In this
solar and wind. work, the relative MAE (rMAE) and relative RMSE (rRMSE) error mea­
The National Institute of Wind Energy (NIWE) had operated few sures are used to compare the data sources as described in the following
wind monitoring stations (WMS) in the state and data is available equations.
sporadically from 1990s (National Institute of Winb). Most of these wind
monitoring stations are situated in the coastal regions of Odisha due to 1∑ n
MAE = |vm,i − vr,i | (1)
the higher wind speeds. Amongst the wind monitoring stations located n i=1

7
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 8. Fitted distributions for MES50m seasonal wind speed.

Fig. 9. Fitted distributions for MER50m seasonal wind speed.

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n RMSE
RMSE = (vm,i − vr,i )2 (2) rRMSE = (4)
n i=1 vm,mean

where vm,i is the measured data, vr,i is the MERRA-2 reanalysis data, and
MAE
rMAE = (3) the number of data points is indicated by n. Finally, the coefficient of
vm,mean
correlation (R) can compare the inherent trends in the time series where

8
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Table 9 Table 10
Distribution Parameters for Monthly Wind speed (MES50m). Distribution Parameters for Monthly Wind speed (MER50m).
Month PD P1 P2 P3 pChi pKS Month PD P1 P2 P3 pChi pKS

Jan Weibull 3.31 4.74 – 0.0043 0.2240 Jan Weibull 2.23 4.41 0.0000 0.0535
Normal 4.27 1.41 – 0.5627 0.6627 Normal 3.92 1.85 0.0024 0.0556
Logistic 4.29 0.80 – 0.6976 0.6429 Logistics 3.89 1.09 0.0000 0.0329
GEV − 0.29 1.44 3.77 0.1037 0.1857 Nakagami 1.11 18.75 0.0000 0.0141
Feb Weibull 2.30 4.84 – 0.5880 0.1475 GEV − 0.25 1.79 3.24 0.0140 0.1954
Normal 4.31 1.96 – 0.3579 0.8202 Feb Weibull 1.97 4.27 0.0530 0.4133
Logistic 4.27 1.12 – 0.0746 0.4671 Gamma 2.99 1.27 0.0001 0.0582
Nakagami 1.19 22.39 – 0.1458 0.0268 Nakagami 0.96 18.37 0.0667 0.4248
GEV − 0.17 1.84 3.53 0.7245 0.4277 GEV − 0.11 1.76 2.93 0.0009 0.1204
Mar Weibull 2.44 6.20 – 0.0002 0.0363 Mar Normal 5.20 2.04 0.0009 0.0950
Normal 5.51 2.40 – 0.0014 0.1024 Logistic 5.24 1.20 0.0000 0.0537
GEV − 0.30 2.40 4.69 0.0301 0.1609 GEV − 0.36 2.10 4.55 0.0071 0.3522
Apr GEV − 0.35 2.78 6.42 0.0167 0.0509 Apr Logistic 7.24 1.17 0.0000 0.0239
June Weibull 2.39 7.68 – 0.0038 0.1675 GEV − 0.37 2.17 6.45 0.0000 0.0102
Gamma 4.40 1.55 – 0.0021 0.1300 May Logistic 8.81 1.52 0.0000 0.0207
Nakagami 1.34 55.55 – 0.0203 0.4362 GEV − 0.43 2.87 7.82 0.0000 0.0528
GEV − 0.13 2.70 5.56 0.0893 0.4699 June Weibull 2.57 7.69 0.3301 0.3794
July Weibull 2.09 6.56 – 0.0518 0.4868 Gamma 4.70 1.46 0.0000 0.0345
Normal 5.83 2.89 – 0.0407 0.1356 Logistic 6.74 1.66 0.0003 0.0836
Nakagami 1.02 42.34 – 0.0254 0.2500 LogLogistic 1.86 0.27 0.0000 0.0112
GEV − 0.18 2.67 4.69 0.3428 0.6598 Nakagami 1.46 54.85 0.1141 0.3712
Aug Weibull 2.20 6.16 – 0.0197 0.3509 GEV − 0.21 2.70 5.74 0.4184 0.2810
Logistics 5.36 1.50 – 0.0000 0.0726 July Normal 7.17 2.70 0.0033 0.3777
Nakagami 1.12 36.68 – 0.0080 0.1016 Logistic 7.17 1.59 0.0000 0.1400
GEV − 0.18 2.40 4.45 0.0561 0.2936 Nakagami 1.75 58.67 0.0002 0.0332
Sep Weibull 2.21 4.66 0.0000 0.0769 GEV − 0.34 2.74 6.30 0.0311 0.2769
Logistics 4.01 1.09 – 0.0000 0.3727 Aug Weibull 2.28 6.93 0.0000 0.0107
LL 1.34 0.30 – 0.0000 0.2010 Normal 6.16 2.82 0.0000 0.0736
Nakagami 1.18 20.94 – 0.0001 0.1176 Logistic 6.09 1.59 0.0000 0.0584
GEV − 0.10 1.71 3.31 0.0033 0.4063 Sep GEV − 0.35 2.10 4.22 0.0000 0.0362
Oct IG 6.74 7.26 – 0.0368 0.0000 Oct Gamma 2.21 2.68 0.0000 0.0352
LL 1.60 0.42 – 0.0020 0.0741 Lognormal 1.54 0.73 0.0003 0.1674
GEV 0.32 2.67 3.94 0.0043 0.0440 LL 1.56 0.40 0.0083 0.6997
Nov Weibull 2.00 7.56 – 0.0000 0.0185 GEV 0.21 2.52 3.86 0.0019 0.9337
Nakagami 1.01 57.04 – 0.0000 0.0197 Nov Weibull 2.65 4.65 0.0000 0.0845
GEV − 0.06 2.90 5.18 0.0000 0.0261 Normal 4.13 1.69 0.0000 0.1497
Dec Logistics 4.28 1.24 – 0.0000 0.0428 Logistic 4.14 0.99 0.0000 0.0837
GEV 0.00 1.88 3.46 0.0000 0.0104 GEV − 0.29 1.69 3.54 0.0000 0.2052
Dec Weibull 2.63 4.22 0.4610 0.7305
Normal 3.75 1.54 0.6030 0.6081
the relationship between the variations in the measured and MERRA-2 Logistic 3.73 0.89 0.0105 0.4368
Nakagami 1.49 16.44 0.0187 0.1196
wind speed are expressed as a number in 0–1 range with higher value
GEV − 0.24 1.49 3.18 0.7775 0.7767
of R signifying high similarity.
∑n
(vm,i − vr,i )
R = i=1√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n 2 (5) parameters. Logistic distribution has mean μ and the scale parameter σ
i=1 vm,i and Log-Logistic distribution has log mean and log scale parameters μ
and σ respectively. The Birnbaum-Saunders (BS) distribution used
extensively in material fatigue and reliability studies has a scale and
3.2. Marginal distributions shape parameters β and γ respectively. The Nakagami distribution is
related to Gamma distribution and has a shape parameter μ and a scale
Since the variation in wind speed is location specific, a number of parameter ω. The GEV distribution has scale parameter k, shape
marginal distributions have been proposed in past research works for parameter σ and a location parameter μ and is therefore a
modelling of wind speed data. The present work takes recourse to ten three-parameter distribution. The shape parameter is representative of
most popular distributions which are summarized in (Mohammadi et al., the tail behaviour of the distribution whereas the scale parameter de­
2017). All the selected distributions, except the Generalized Extreme termines the spread of the distribution. The location parameter μ de­
Value (GEV) are two-parameter distributions. For ready reference, the termines the location or shift of the distribution.
parameters of the selected distributions are provided in Table 1 (Samal, It is instructive to provide an overview of advantages and disad­
2021) and are briefly described below. Weibull distribution is the most vantages of the wind speed probability distributions described in
widely used and in fact the most preferred distribution for modelling Table 1. Weibull PD is a two parameter distribution and its popularity is
wind speed. It is a two-parameter distribution having shape parameter k due to the fact that the scale factor is directly related to the mean wind
and scale parameter c. The shape parameter k indicates the shape of the speed and can provide a glimpse on the energy content whereas the scale
distribution and the scale parameter is indicative of the amount of en­ factor can be directly used to determine turbine parameters (Samal and
ergy available in the wind. Gamma distribution also has shape and scale Tripathy, 2017). The primary disadvantage is that it is not suitable for
parameters represented by a and b respectively. Normal distribution in modelling low wind speeds (below 2 m/s) (Ayik et al., 2021). The
its exact form cannot model wind resource because of non-negative advantage of Gamma PD is that it is the best PD wherever Weibull is a
nature of wind speed. However, if the normal distribution is truncated close approximation (Feijóo and Villanueva, 2016). However, the very
below zero, known as truncated Normal (TN), it can be used as a wind similarity of Gamma with Weibull can be considered to be its main
speed distribution parametrized using mean μ and standard deviation σ . disadvantage because wherever Gamma is suitable Weibull PD is
Lognormal distribution are log mean μ and log standard deviation σ generally applicable. Both the LN and IG PD can represent wind speeds
parameters; Inverse Gaussian distribution has mean μ and shape λ

9
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 10. Best Fitted distribution for MES50m Jan–Apr.

Fig. 11. Best Fitted distribution for MES50m May–Aug.

where Weibull and Gamma are not appropriate, however their disad­ 2011). The B–S PD is an asymmetric and positively skewed distribution
vantage is that the parameters do not provide much intuition about the which has been found suitable for modelling positive data, however, it
wind resource of the location. The disadvantage of Normal PD is that it has been found suitable for modelling wind speed only for very few
assumes negative values whereas wind speed is non-negative requiring locations (Mohammadi et al., 2017). The disadvantage of Logistic, LL
the use of Truncated Normal whereas its advantage is that it provides a and Nakagami PDs is that they have been shown to model wind speed
better fit in some cases where Weibull PD is also applicable (Chang, only for fewer locations and do not provide much intuition on the nature

10
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 12. Best Fitted distribution for MES50m Sep–Dec.

Fig. 13. Best Fitted distribution for MER50m Jan–Apr.

of wind resource. The advantage of Logistic PD is its resemblance to distributions combines three simpler distributions where the tail of Type
Normal PD but with higher kurtosis whereas the advantage of LL PD is I distributions decrease exponentially, that of Type II decrease as a
its wide use in hydrological studies which is similar to wind resource polynomial and the tails of Type III distributions are finite. GEV PD has
analysis. The advantages of Nakagami distribution is its applicability in the disadvantage of being a three parameter distribution and therefore
a wide range of fields which indicates its potential for universal appli­ clumsy to work with. However, it has the advantage of being applicable
cability though more studies are needed (Alavi et al., 2016). GEV to wind speed of regions where none of the previously mentioned

11
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 14. Best Fitted distribution for MER50m May–Aug.

Fig. 15. Best Fitted distribution for MER50m Sep–Dec.

distributions are suitable (Samal and Tripathy, 2018; Samal, 2021). and energy pattern factor methods. The above methods have largely
been investigated in the context of Weibull distribution (Zhang et al.,
2020). Different studies have concluded in favour of one or the other
3.3. Parameter estimation parameter estimation method. For example, MLE method has been
found to be most suitable (Shoaib et al., 2019). In fact, all these methods
Some of the popular parameter estimation techniques are least come with certain merits and demerits. Alternatively, different software
squares, maximum likelhood (MLE), modified MLE, method of moments

12
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

Fig. 16. GOF Statistics for Weibull distribution across months.

threshold value. Similarly, the fit is accepted if the p-Value higher than a
Table 11
threshold, which is either 0.01 or 0.05 corresponding to 99% or 95%
Weibull parameters and WPD.
significance levels.
Period MES50m MER50m Difference in Selecting a parameter estimation method and choosing a GOF test
WPD (%)
k c WPD k c WPD are non-trivial tasks, however, their significance decreases as the num­
(W/m2) (W/m2) ber of data points increase. The focus of this work is on comparison of
Jan 3.31 4.74 63 2.23 4.41 63 − 0.5 probability distributions rather than comparing the effectiveness of
Feb 2.30 4.84 81 1.97 4.27 65 20.8 parameter estimation methods. Therefore, the inbuilt functions of
Mar 2.44 6.20 163 2.80 5.84 126 23.1 MATLAB Statistics Toolbox are conveniently selected for parameter
Apr 3.02 8.13 329 4.05 7.82 269 18.2
May 3.06 9.43 510 3.72 9.48 488 4.4
estimation and performing GOF tests. Specifically, the vert fit distvert
Jun 2.39 7.68 316 2.57 7.69 302 4.4 function is used for estimating parameters of the wind speed distribu­
Jul 2.09 6.56 220 2.93 8.04 321 − 45.9 tions. The chi2 and K–S GOF tests are performed by the functions vert
Aug 2.20 6.16 174 2.28 6.93 240 − 38.3 chi2gofvert and vert kstestvert respectively. The threshold p-value is
Sep 2.21 4.66 75 2.53 5.50 111 − 48.2
selected to be 0.01 and hence the distribution fitting procedure must
Oct 1.25 7.31 715 1.49 6.62 361 49.4
Nov 2.00 7.56 350 2.65 4.65 65 81.4 yield a p-value greater than 0.01 for acceptance.
Dec 1.92 5.10 113 2.63 4.22 49 56.5
Summer 2.53 7.96 337 2.85 7.80 297 11.9
Rainy 2.17 6.79 235 2.55 7.56 288 − 22.4 3.4. Wind power density
Autumn 1.46 6.53 360 1.75 5.63 171 52.5
Winter 2.26 4.91 86 2.24 4.31 59 32.0
Wind turbine power curves are the most convenient and hence
Year 1.80 6.57 262 2.01 6.36 209 20.2
widely used for estimation of power output of a wind turbine. However,
at resource assessment state, when the characteristics of the wind
packages such as MATLAB, Python and R have also several powerful resource is not yet determined, an effective measure of wind power can
inbuilt functions available for parameter estimation. More recently, be obtained by wind power density (WPD). WPD can be obtained using
metaheuristic optimization techniques have also been utilized for either using Weibull parameters using (6) or directly from the time series
parameter estimation (Guedes et al., 2020; Alrashidi et al., 2020). using (7) (Dabbaghiyan et al., 2016).
Once the parameters are estimated model diagnostics can be per­ ∫∞ ( )
1 1 3
formed by using goodness-of-fit (GOF) tests. Chi Square test (chi2) and WPDw = ρ V 3 f (V)dV = ρc3 Γ 1 + (6)
2 0 2 k
Kolmogorov-Smirnov (K–S) tests are widely used and most reliable GOF
tests. Other fitness measures include R2 coefficient, Bayesian informa­
where f(V) is the Weibull PDF and ρ is the air density (1.225 kg/m3
tion criterion (BIC) and Akaike information criterion (AIC) or a single
referring to standard atmospheric conditions).
measure combining them (Miao et al., 2019). It is however quite per­
plexing to see the attempt to use a large number of GOF metrics, 1 ∑N
Vj 3
WPDav = ρ (7)
sometimes as high as ten, for evaluating the quality of fit whereas many 2 j=1 N
a times one or two metrics serve the same purpose. The GOF tests can be
interpreted by means of the fitness statistics or the p-Value. The lower where N represents the number of observations in the measured data.
the fitness statistics, the better is the model fit. A particular distribution More detailed classification methods such as wind energy factors are
is accepted as a fit for the measured data only if GOF statistic are below a built upon the concepts of WPD (Zheng et al., 2019).

13
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

4. Results and discussion correlation is in the month of January at − 2.52% for MES20m and in
Nov at 11% for MES50m. The correlation for DA data is above 50% for
Initially, descriptive statistics of the measured and MERRA-2 data are most of the months. A peculiar observation is that in case of measured
compared. Thereafter, comparisons between hourly wind speed for year, data at different heights in November, correlation is 37% for MES50m
months and seasons are performed. This essentially includes obtaining and 82% for MES20m.
error measures, coefficient of correlation and fitting of probability dis­
tributions to the wind resource data. Finally, wind power density is 4.3. Comparison of probability distributions
computed using the data from both the sources.
4.3.1. Yearly distributions
4.1. Statistical analysis All the ten selected marginal distributions are fitted to the MES50m
and MER50m yearly wind speed. It is found that none of the distribu­
The measures of descriptive statistics for the MES50m and MER50m tions could be fitted to the MES50m. For the MER50m yearly data, only
are provided in Table 2 and Table 3 respectively. It can be observed from the K–S test for Nakagami and Weibull are successful. The Chi2 and K–S
the above tables that mean and CV are quite similar for both the data statistics for the all the distributions are plotted in Fig. 5. It is inferred
sources, however, there are large differences in the skewness and kur­ that the Weibull, Gamma, and Nakagami distributions are the best three
tosis values. Except for Rainy season, the mean of MES50m is higher distribution for MES50m whereas Weibull, Nakagami and GEV are the
compared to MER50m mean. Similarly, the CV can be seen to be nearly best three distributions for MER50m. It is also observed that the chi2Stat
same except for Autumn season. The skewness and kurtosis values are for MER50m is much higher than that of MES50m showing a poor fit for
quite similar for Summer and Rainy seasons whereas they differ widely MER50M. The three lowest K–S statistics are for Weibull, Gamma and
for Autumn and Winter seasons. Fig. 1 compares mean and CV of GEV for MES50m and Weibull, Nakagami and GEV for MER50m. The
MES50m and MER50m. It can be observed from the above figure that as histogram of the wind speed superimposed with three best fitted dis­
compared to MES50m, MER50m mean values are same for May and tributions are shown in Figs. 6 and 7 for measured and MERRA-2 wind
June, higher in July, Aug, Sep and lower for the remaining months. The speed respectively. Sharp differences can be visualized between the
difference between the mean wind speed is very high (6.71 m/s 4.13 m/ histograms of MER50m and MES50m data. The parameters of the best fit
s) for November. Except for Jan and Feb, the CV of the MER50m is lower distributions are provided in Table 6.
compared to MER50m. Plots of skewness and kurtosis are provided in
Fig. 2 which shows that skewness of MES50m for Jan, Mar and MER50m 4.3.2. Seasonal distributions
for July, Sep and Nov are nearly zero indicating normality of the data. The marginal distributions are fitted to the hourly average wind
The MES50m for Jan, Apr, May; MER50m for Mar, Apr, May have speeds for all seasons. For MES50m, no marginal distribution is found
negative skewness and the remaining months have positive skewness. It suitable for Summer wind resource; Normal, Nakagami and GEV dis­
can also be observed that except for Oct and Dec, the skewness of both tributions are successful fits to the data for Rainy season; Lognormal and
the data sources are nearly the same. A comparison of kurtosis of the Loglogistic distributions fitted the data for Autumn season and only
data sources shows that except for Oct, Nov and Dec the kurtosis of the Logistics distribution could be fitted to Winter data. For MER50m, the
sources are approximately equal. Especially, MES50m for Oct, Nov, Dec fitted distributions are Normal and GEV for summer; Weibull, Normal
and MER50m for Aug, Oct have positive excess kurtosis showing the and GEV for Rainy; Weibull, Nakagami and GEV for Winter whereas
platykurtic nature of the data. The MES50m for Oct has very high kur­ none could be fitted to Autumn data. The parameters of the fitted and
tosis of 32.70 as compared to kurtosis of 6.9 in MER50m for Oct. The the best fit distributions are provided in Table 7 and Table 8 for MES50m
MES50m for Jan, Feb, Jun and MER50m for Apr, May have nearly zero and MER50m respectively. The most suitable distributions with the least
excess kurtosis showing the mesokurtic nature of the data. MES50m and chi2Stat is plotted superimposed on the histogram of the data in Fig. 8
MER50m for remaining months have negative excess kurtosis showing and Fig. 9 for MES50m and MER50m respectively. It can inferred that it
the leptokurtic nature of the data for these months. is the GEV PD that is the most fitted distribution for the seasonal data.

4.2. Error measures and correlation 4.3.3. Monthly distributions


The results of fitting MES50m monthly HA data is provided in
The rMAE and rRMSE error measures for yearly and seasonal data Table 9. The above table contains the distributions that are successful in
are provided in Table 4. Very high error measures of 33.57% and least one of the GOF tests. As described earlier, the highest p-Value
49.22% can be observed for yearly average data. The error measures are determines the best fit. The best fit distribution with respect to p-Value
the highest for Autumn followed by Winter. Due to averaging effect, the of chi2 test for the months are as follows: Logistic for Jan; GEV for Feb,
error measures for daily average data are much less compared to HA Mar and Apr, June, July, Aug, Sep, Nov; Inverse Gaussian for Oct and
data. The variations of error measures across months is provided in Logistic for Dec. As with seasonal data, the most widely fitted distribu­
Fig. 3. It can be observed from the above figure that Nov has the highest tion for the monthly data is the GEV PD. None of the distributions could
rMAE of 49% followed by February at 44%. The rRMSE error is highest be fitted to May data. The results of fitting MER50m data is provided in
for October at 73% followed by February at 55%.In the next step, the Table 10. It is inferred that a large number of distributions could be
coefficient of correlation between the data sources are obtained. Table 5 fitted to all the months except Apr, May and Sep. The best fitted dis­
contains the correlation between MES50m and MER50m and also be­ tributions with respect to chi2 p-Value are GEV for Jan, Mar, May, June,
tween MES20m and MER10m. It can be observed that the correlation for July, Sep, Nov; Nakagami for Feb; Logistic for Apr; Normal for Aug and
the HA and DA data are satisfactory at 69% and 85% respectively for Dec; LL for Oct. It is worthwhile to note that more than one distribution
MES20m and 60% and 76% for MES50m. The lowest correlation is for are found suitable for a number of months and the GOF test resulted in
Winter season at both measurement heights. MES20m for both Summer very high p-values as compared to the threshold. Fig. 10, Fig. 11, Fig. 12
and Autumn seasons show very high correlation values whereas it is for depicts the best fitted distribution superimposed on monthly data for
Summer and Rainy season that MES50m has high correlation. The co­ MES50m. whereas those for MER50m are depicted in Figs. 13–15.
efficient of correlation for all the months are depicted in Fig. 4. It can be
observed that in the month of March and October the HA correlation for 4.4. Comparison of Weibull parameters and WPD
MES20m is greater than that of MES50m. For HA data, the highest
monthly correlation is obtained for October at 86% and 67% respec­ In commercial wind power development, Weibull distribution is the
tively for MES20m and MES50m respectively. The lowest value of most preferred and widespread for performing power and energy anal

14
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

ysis irrespective of its applicability to modelling the wind speed of the statistics, error measures, wind speed probability distributions, and
concerned region. Therefore, at this stage, it seems appropriate to wind power density. Most of the analysis is done for wind speed at 50 m
discuss the applicability of Weibull distribution to the selected region. In measurement height. Correlation measures between 20 m wind speed
the earlier sections, it is observed that although found unsuitable for and MERRA-2 10 m wind speed are also obtained. It is found that the
yearly wind speed, Weibull PD fits the wind speed of eight months mean and variance of wind resource from both the sources can be
namely Jan, Feb, Mar, June, July, Aug, Sep and Nov. Weibull PD is also considered to be consistent for yearly and seasonal data. Large differ­
observed to be suitable for modelling wind resource in rainy season. ences in descriptive statistical measures are observed for monthly data.
Similarly, with MERRA-2, Weibull PD fitted the wind speeds of seven The yearly correlation is found to be satisfactory though the correlation
months namely, Jan, Feb, June, July, Aug, Nov and Dec and also for values for months and seasons are lower. The error measures obtained
Rainy and Winter seasons. The GOF statistics of Weibull distributions reveal that hourly wind speed obtained from the two sources differ
shown in Fig. 16 for all the months ascertains its applicability. There­ widely. Probability distribution fitting indicates that Weibull and GEV to
fore, it is relevant to list the Weibull parameters of all the months and be the most common distributions for the data from both the sources. In
seasons which can be used for computation of WPD. The shape and scale fact, barring Birnbaum Saunders distribution, all the distributions are
parameters and the corresponding WPD for the complete year, seasons fitted at least once. The Weibull parameters for the year and seasons
and months are listed in Table 11, where the last column shows the obtained from both sources are found to be in agreement. However, for
difference in WPD values based on (8). monthly wind speed, wind power density computed from both the
sources differ widely. Therefore, it can be concluded that the measured
WPDMES50m − WPDMER50m
× 100 (8) and MERRA-2 data can be assumed to be in agreement for longer du­
WPDMES50m
rations such as a year but not so for seasonal or monthly durations. This
As inferred from Table 11, the yearly WPD shows good wind power signifies that the reanalysis data can only be used for long term planning;
potential based on the classification in (Zheng et al., 2016). The differ­ it is however unsuitable for short term studies.
ence in yearly WPD computed from both sources is found to be 20.2%. The major limitation of this study is the availability of a single year of
Higher differences in monthly WPD values computed using measured measured wind data for the selected region. however, this is the case
and MERRA-2 data are also observed. The highest discrepancy occurs in with most of the wind monitoring stations due to associated logistics and
the month of Nov where the WPD computed using MES50m is found to cost. In fact, this work aspires to address this very limitation by showing
be 81.4% higher than that computed using MER50m. Such a high dif­ that the similarity between reanalysis and short-term measured wind
ferences in WPD warrants caution while using MERRA-2 data for energy resource can be leveraged to simulate multi-year wind speed data for
planning purposes especially for shorter time spans. wind power potential studies. Another limitation of this work is that it
stops at comparing the wind power density which, though useful, is a
4.5. Discussion crude measure of wind energy resource. This gaps can be addressed in a
separate future work where a comparison of economic evaluation of
The mean yearly wind speed from both measured and MERRA-2 wind power projects in the region based on measured and reanalysis
sources are seen to be very close at 5.85 m/s and 5.64 m/s respec­ sources can be performed. A third limitation, though not a prominent
tively, that is, a difference of about 3.6%. The differences between the one, is the use of commercial toolbox functions for parameter estimation
seasonal wind speeds are 1.8%, 11.4%, 15%, 12.6% for Summer, Rainy, which can be alleviated if open source software can be used. Thus
Winter and Autumn seasons respectively. In terms of monthly mean another prospective research direction can be the use of metaheuristic
values, the closest resemblance occurs in June (6.81 and 6.84 m/s) and techniques using open source software for parameter estimation of wind
the largest difference occurs in Nov (38%). The differences are below speed distributions.
10% for five months and below 20% for ten months. The values of co­
efficient of variation are in agreement for yearly and seasonal data, Data availability
however, they vary widely across months. Substantial variations in
skewness and kurtosis are observed for yearly, seasonal and monthly The information regarding the reanalysis data is available at (GrADS
data signifying differences in shape of the frequency distribution. The Data Server). The mast measured data that support the findings of this
error measures reveal large difference between the measured and study are proprietary of the National Institute of Wind Energy, Govt. of
MERRA-2 data. Values of yearly correlation for hourly and daily average India, and are procured under institutional use terms.
data at (69%, 86%) for 20 m and (60%,76%) for 50 m data are
acceptable. However, the correlation varies across seasons and months. CRediT authorship contribution statement
For example, for the 50 m data it varies from 22% in Winter to about
70% in Summer and from 11% in Nov to 76% in June. Probabilistic Rajat Kanti Samal: sole author performing all tasks related to the
modelling attempts reveal that GEV is the most commonly fitted dis­ manuscript including, Conceptualization, Funding acquisition, Meth­
tribution for the data both the sources. Weibull and Nakagami distri­ odology, Software, Validation, Writing – original draft.
butions are fitted to the yearly MERRA-2 data, however, no distribution
is fitted to the yearly measured wind speed. Seasonal data from both the Declaration of competing interest
sources could be fitted with only a few distributions. A large number of
distributions could be fitted to the monthly wind speed from both the The authors declare that they have no known competing financial
sources, differing only in values of parameters. It is also found that from interests or personal relationships that could have appeared to influence
among the selected distributions, the Birnbaum-Sanuders distribution the work reported in this paper.
could not be fitted at all to wind speed from both the sources and can be
considered unsuitable for representing wind speed uncertainty in this Acknowledgement(s)
region. The Weibull parameters are seen to be different as computed
from measured and MERRA-2 data and so is wind power density. This research is financially supported by the research grant under
Collaborative Research and Innovation Scheme of Veer Surendra Sai
5. Conclusion University of Technology (Ref. No. VSSUT/TEQIP/34/2020).

In this work, a comparison is performed between commercially


measured wind resource and the NASA MERRA-2 in terms of descriptive

15
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

References Li, Jiale, Yu, Xiong (Bill), 2018. Onshore and offshore wind energy potential assessment
near Lake Erie shoreline: a spatial and temporal analysis. Energy 147, 1092–1107.
https://doi.org/10.1016/j.energy.2018.01.118. ISSN 0360-5442.
Adedeji, Paul A., Akinlabi, Stephen A., Madushele, Nkosinathi, Olatunji, Obafemi O.,
Li, Yi, Wu, Xiao-Peng, Li, Qiu-Sheng, Tee, Kong Fah, 2018. Assessment of onshore wind
2020. Neuro-fuzzy resource forecast in site suitability assessment for wind and solar
energy potential under different geographical climate conditions in China. Energy
energy: a mini review. J. Clean. Prod. 269, 122104. https://doi.org/10.1016/j.
152, 498–511. https://doi.org/10.1016/j.energy.2018.03.172. ISSN 0360-5442.
jclepro.2020.122104. ISSN 0959-6526.
Miao, Shuwei, Gu, Yingzhong, Li, Dan, Han, Li, 2019. Determining suitable region wind
Alavi, Omid, Mohammadi, Kasra, Ali, Mostafaeipour, 2016. Evaluating the suitability of
speed probability distribution using optimal score-radar map. Energy Convers.
wind speed probability distribution models: a case of study of east and southeast
Manag. 183, 590–603. https://doi.org/10.1016/j.enconman.2019.01.001. ISSN
parts f Iran. Energy Convers. Manag. 119, 101–108. https://doi.org/10.1016/j.
0196-8904.
enconman.2016.04.039. ISSN 0196-8904.
Modern-Era Retrospective analysis for Research and Applications, Version 2, Accessed
Alrashidi, Musaed, Rahman, Saifur, Pipattanasomporn, Manisa, 2020. Metaheuristic
June 20, 2018. http://gmao.gsfc.nasa.gov.j. doi: 10.5067/VJAFPLI1CSIV..
optimization algorithms to estimate statistical distribution parameters for
Mohammadi, Kasra, Alavi, Omid, McGowan, Jon G., 2017. Use of Birnbaum-Saunders
characterizing wind speeds. Renew. Energy 149, 664–681. https://doi.org/10.1016/
distribution for estimating wind speed and wind power probability distributions: a
j.renene.2019.12.048. ISSN 0960-1481.
review. Energy Convers. Manag. 143, 109–122. https://doi.org/10.1016/j.
Ayik, A., Ijumba, N., Kabiri, C., Goffin, P., 2021. Preliminary wind resource assessment in
enconman.2017.03.083. ISSN 0196-8904.
South Sudan using reanalysis data and statistical methods. Renew. Sustain. Energy
Murthy, K.S.R., Rahi, O.P., 2017. A comprehensive review of wind resource assessment.
Rev. 138, 110621. https://doi.org/10.1016/j.rser.2020.110621. ISSN 1364-0321.
Renew. Sustain. Energy Rev. 72, 1320–1342. https://doi.org/10.1016/j.
Bina, Saeid Mohammadzadeh, Jalilinasrabady, Saeid, Fujii, Hikari, Farabi-Asl, Hadi,
rser.2016.10.038. ISSN 1364-0321.
2018. A comprehensive approach for wind power plant potential assessment,
National Institute of Wind Energy. Wind power potential at 100 m agl. Online. Available:
application to northwestern Iran. Energy 164, 344–358. https://doi.org/10.1016/j.
vert https://niwe.res.in/department_wra_100m%20agl.phpvert.
energy.2018.08.211. ISSN 0360-5442.
National Institute of Wind Energy. List of wind monitoring stations. Online. Available:
Campisi-Pinto, Salvatore, Gianchandani, Kaushal, Yosef, Ashkenazy, 2020. Statistical
vert https://niwe.res.in/department_wra_lwms.phpvert.
tests for the distribution of surface wind and current speeds across the globe. Renew.
Odisha Renewable Energy Development Agency. Online. Available: http://oredaodisha.
Energy 149, 861–876. https://doi.org/10.1016/j.renene.2019.12.041. ISSN 0960-
com/achivements.htm..
1481.
Olaofe, Z.O., 2018. Review of energy systems deployment and development of offshore
Carvalho, D., Rocha, A., Gómez-Gesteira, M., Silva Santos, C., 2014. WRF wind
wind energy resource map at the coastal regions of Africa. Energy 161, 1096–1114.
simulation and wind energy production estimates forced by different reanalyses:
https://doi.org/10.1016/j.energy.2018.07.185. ISSN 0360-5442.
comparison with observed data for Portugal. Appl. Energy 117, 116–126. https://
Olauson, Jon, Bergkvist, Mikael, 2015. Modelling the Swedish wind power production
doi.org/10.1016/j.apenergy.2013.12.001. ISSN 0306-2619.
using MERRA reanalysis data. Renew. Energy 76, 717–725. https://doi.org/
Chancham, Chana, Waewsak, Jompob, Gagnon, Yves, 2017. Offshore wind resource
10.1016/j.renene.2014.11.085. ISSN 0960-1481.
assessment and wind power plant optimization in the Gulf of Thailand. Energy 139,
Ren, Guorui, Wan, Jie, Liu, Jinfu, Yu, Daren, 2019. Characterization of wind resource in
706–731. https://doi.org/10.1016/j.energy.2017.08.026. ISSN 0360-5442.
China from a new perspective. Energy 167, 994–1010. https://doi.org/10.1016/j.
Chandel, S.S., Ramasamy, P., Murthy, K.S.R., 2014a. Wind power potential assessment of
energy.2018.11.032. ISSN 0360-5442.
12 locations in western Himalayan region of India. Renew. Sustain. Energy Rev. 39,
Rienecker, M.M., Suarez, M.J., Gelaro, R., Todling, R., Liu, J. BacmeisterE.,
530–545. https://doi.org/10.1016/j.rser.2014.07.050. ISSN 1364-0321.
Bosilovich, M.G., Schubert, S.D., Takacs, L., Kim, G., Bloom, S., Chen, J., Collins, D.,
Chandel, S.S., Murthy, K.S.R., Ramasamy, P., 2014b. Wind resource assessment for
Conaty, A., da Silva, A., Gu, W., Joiner, J., Koster, R.D., Lucchesi, R., Molod, A.,
decentralised power generation: case study of a complex hilly terrain in western
Owens, T., Pawson, S., Pegion, P., Redder, C.R., Reichle, R., Robertson, F.R.,
Himalayan region. Sustain. Energy Technol. Assess. 8, 18–33. https://doi.org/
Ruddick, A.G., Sienkiewicz, M., Woollen, J., 2011. MERRA: NASA’s Modern-Era
10.1016/j.seta.2014.06.005. ISSN 2213-1388.
retrospective analysis for research and applications. J. Clim. 24, 3624–3648. https://
Chang, Tian Pau, 2011. Estimation of wind energy potential using different probability
doi.org/10.1175/JCLI-D-11-00015.1.
density functions. Appl. Energy 88 (Issue 5), 1848–1856. https://doi.org/10.1016/j.
Samal, Rajat Kanti, 2021. Probabilistic modelling of 80 m mast measured wind resource:
apenergy.2010.11.010. ISSN 0306-2619.
a case study. Energy Sources, Part A Recovery, Util. Environ. Eff. https://doi.org/
Cook, Nicholas J., 2019. The OEN mixture model for the joint distribution of wind speed
10.1080/15567036.2021.1910753.
and direction: a globally applicable model with physical justification. Energy
Samal, R.K., Tripathy, M., 2017. Wind resource assessment and energy analysis for wind
Convers. Manag. 191, 141–158. https://doi.org/10.1016/j.enconman.2019.04.015.
energy projects. In: 2017 6th International Conference on Computer Applications in
ISSN 0196-8904.
Electrical Engineering-Recent Advances (CERA), pp. 128–133. https://doi.org/
Dabar, Omar Assowe, Awaleh, Mohamed Osman, Kirk-Davidoff, Daniel, Olauson, Jon,
10.1109/CERA.2017.8343314. Roorkee.
Söder, Lennart, Awaleh, Said Ismael, 2019. Wind resource assessment and economic
Samal, Rajat Kanti, Tripathy, Manish, 2018. Estimating wind speed probability
analysis for electricity generation in three locations of the Republic of Djibouti.
distribution based on measured data at Burla in Odisha, India. Energy Sources, Part
Energy 185, 884–894. https://doi.org/10.1016/j.energy.2019.07.107. ISSN 0360-
A Recovery, Util. Environ. Eff. https://doi.org/10.1080/15567036.2018.1521888.
5442.
Samal, Rajat Kanti, Tripathy, M., 2019. A novel distance metric for evaluating impact of
Dabbaghiyan, A., Fazelpour, Farivar, Abnavi, Mohhamadreza Dehghan, Rosen, Marc A.,
wind integration on power systems. Renew. Energy 140, 722–736. https://doi.org/
2016. Evaluation of wind energy potential in province of Bushehr, Iran. Renew.
10.1016/j.renene.2019.03.094. ISSN 0960-1481.
Sustain. Energy Rev. 55, 455–466. https://doi.org/10.1016/j.rser.2015.10.148,
Sharma, Pramod Kumar, Warudkar, Vilas, Ahmed, Siraj, 2019. Application of lidar and
1364-0321.
measure correlate predict method in offshore wind resource assessments. J. Clean.
Duran, Pablo, Meibner, Catherine, Casso, Pau, 2020. A New Meso-Microscale Coupled
Prod. 215, 534–543. https://doi.org/10.1016/j.jclepro.2018.12.267. ISSN 0959-
Modelling Framework for Wind Resource Assessment: A Validation Study.
6526.
Renewable Energy. https://doi.org/10.1016/j.renene.2020.06.074. ISSN 0960-
Shoaib, Muhammad, Siddiqui, Imran, Rehman, Shafiqur, Khan, Shamim, Luai, M., 2019.
1481.
Alhems, Assessment of wind energy potential using wind energy conversion system.
Elsner, Paul, 2019. Continental-scale assessment of the African offshore wind energy
J. Clean. Prod. 216, 346–360. https://doi.org/10.1016/j.jclepro.2019.01.128. ISSN
potential: spatial analysis of an under-appreciated renewable energy resource.
0959-6526.
Renew. Sustain. Energy Rev. 104, 394–407. https://doi.org/10.1016/j.
Staffell, Iain, Green, Richard, 2014. How does wind farm performance decline with age?
rser.2019.01.034. ISSN 1364-0321.
Renew. Energy 66, 775–786. https://doi.org/10.1016/j.renene.2013.10.041. ISSN
Feijóo, Andrés, Villanueva, Daniel, 2016. Assessing wind speed simulation methods,
0960-1481.
Renewable and Sustainable Energy Reviews 56, 473–483. https://doi.org/10.1016/
Valsaraj, P., Alex Thumba, Drisya, Asokan, K., Kumar, K. Satheesh, 2020. Symbolic
j.rser.2015.11.094. ISSN 1364-0321.
regression-based improved method for wind speed extrapolation from lower to
GrADS Data Server - Info for M2T1NXSLV. Online. Available: vert https://goldsmr4.ges
higher altitudes for wind energy applications. Appl. Energy 260, 114270. https://
disc.eosdis.nasa.gov/dods/M2T1NXSLV.infovert.
doi.org/10.1016/j.apenergy.2019.114270. ISSN 0306-2619.
Guedes, Kevin S., de Andrade, Carla F., Rocha, Paulo A.C., Mangueira, Rivanilso dos S.,
Wang, Jianzhou, Huang, Xiaojia, Li, Qiwei, Ma, Xuejiao, 2018. Comparison of seven
de Moura, Elineudo P., 2020. Performance analysis of metaheuristic optimization
methods for determining the optimal statistical distribution parameters: a case study
algorithms in estimating the parameters of several wind speed distributions. Appl.
of wind energy assessment in the large-scale wind farms of China. Energy 164,
Energy 268, 114952. https://doi.org/10.1016/j.apenergy.2020.114952. ISSN 0306-
432–448. https://doi.org/10.1016/j.energy.2018.08.201. ISSN 0360-5442.
2619.
Zare Oskouei, Morteza, Mohammadi-Ivatloo, Behnam, Abapour, Mehdi, Mahmood
Hernández, Cristina Vázquez, González, Javier Serrano, Fernández-Blanco, Ricardo,
Shafiee, Anvari-Moghaddam, Amjad, 2021. Techno-economic and environmental
2019. New method to assess the long-term role of wind energy generation in
assessment of the coordinated operation of regional grid-connected energy hubs
reduction of CO2 emissions – case study of the European Union. J. Clean. Prod. 207,
considering high penetration of wind power. Part 1 J. Clean. Prod. 280. https://doi.
1099–1111. https://doi.org/10.1016/j.jclepro.2018.09.249. ISSN 0959-6526.
org/10.1016/j.jclepro.2020.124275, 124275, ISSN 0959-6526.
Jung, Christopher, Schindler, Dirk, 2018. Sensitivity analysis of the system of wind speed
Zhang, Jingyu, Zhang, Mingjin, Li, Yongle, Qin, Jingxi, Wei, Kai, Song, Lili, 2020.
distributions. Energy Convers. Manag. 177, 376–384. https://doi.org/10.1016/j.
Analysis of wind characteristics and wind energy potential in complex mountainous
enconman.2018.09.078. ISSN 0196-8904.
Kubik, M.L., Brayshaw, D.J., Coker, P.J., Barlow, J.F., 2013. Exploring the role of
reanalysis data in simulating regional wind generation variability over Northern
Ireland. Renew. Energy 57, 558–561. https://doi.org/10.1016/j.
renene.2013.02.012. ISSN 0960-1481.

16
R.K. Samal Journal of Cleaner Production 313 (2021) 127933

region in southwest China. J. Clean. Prod. 274, 123036. https://doi.org/10.1016/j. Zheng, Chong-wei, Li, Chong-yin, Xu, Jian-jun, 2019. Micro-scale classification of
jclepro.2020.123036. ISSN 0959-6526. offshore wind energy resource - a case study of the New Zealand. J. Clean. Prod. 226,
Zheng, Chong Wei, Li, Chong Yin, Pan, Jing, Liu, Ming Yang, Xia, Lin Lin, 2016. An 133–141. https://doi.org/10.1016/j.jclepro.2019.04.082. ISSN 0959-6526.
overview of global ocean wind energy resource evaluations. Renew. Sustain. Energy
Rev. 53, 1240–1251. https://doi.org/10.1016/j.rser.2015.09.063. ISSN 1364-0321.

17

You might also like