Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Sustainable Energy Technologies and Assessments 52 (2022) 101995

Contents lists available at ScienceDirect

Sustainable Energy Technologies and Assessments


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

Original article

A multi-scale model for day-ahead wind speed forecasting: A case study of


the Houhoku wind farm, Japan
Yuzhang Che a, b, Andrés A. Salazar b, Siyue Peng a, Jiafeng Zheng a, Yangruixue Chen a,
Liang Yuan c, *
a
College of Atmospheric Science, Chengdu University of Information Technology, Chengdu 610225, China
b
Department of Mechanical Engineering, Tokyo Institute of Technology, Tokyo 152-8550, Japan
c
Chengdu Plain Urban Meteorology and Environment Observation and Research Station of Sichuan Province, Chengdu University of Information Technology, Chengdu
610225, China

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

Keywords: Accurate day-ahead wind prediction from numerical weather prediction (NWP) models is crucial for integrating
Multi-scale wind power into existing grids and wind farm arrangements. Most wind farms in Japan are scattered in
WRF mountainous areas susceptible to strong winds and wind patterns that are highly fluctuating and cannot be
Computational Fluid Dynamics
captured well with mesoscale NWP models. Aiming to provide reliable day-ahead forecasts in rugged zones with
Wind prediction
Coupling
complex topography, we introduce a multi-scale model based on coupling the Weather Research Forecasting
(WRF) model and the Open Source Field Operation and Manipulation (OpenFOAM) Computational Fluid Dy­
namics model. The WRF’s predictions are used as initial/boundary conditions for a microscale simulation using
OpenFOAM. This coupled model can overcome the shortcomings when employing each component indepen­
dently. Numerical simulations and validations are carried out at a wind farm in Houhoku, Japan, during summer
and winter. The multi-scale model results show a considerable reduction in the bias, mean absolute error, and
root mean square error compared to the WRF-alone forecasts. It also has a superior skill against most post-
processing algorithms or artificial intelligence techniques to reduce random forecasting errors. In addition,
analysis of computational cost demonstrates the promising capability of this multi-scale model being a routine
tool.

Introduction such as Netherlands. The difficulty of simulating and predicting low-


level wind might be one of the factors that limit the share of wind en­
Wind energy has the tremendous potential to be the solution to the ergy to total power generation in Japan, which as of 2019 is only 0.76%.
shortage of fuel and greenhouse-gas-induced climate change because of On the other hand, fossil fuels, including coal, LNG, and oil have a 75%
its clean, renewable, and sustainable characteristics. Driven by decades [2] share. Although this metric has decreased from 78% the previous
of scientific exploration and technology development, it has already year, it is still worryingly high. Concerning the world’s environmental
played a crucial role as a mainstream energy resource nowadays. Its situation and the reduction of nuclear power supply, the need for
global cumulative installed capacity is increasing rapidly in recent years carbon-free and pollution-free wind energy is more urgent than ever
with 651 GW installed as of 2019. It is expected that over 355 GW of new before. Wind power generation is, however, different from other tradi­
capacity will be added between 2020–2024 [1]. However, the utilization tional sources in that its availability depends on the atmospheric con­
efficiencies of wind remain low, and the abandoned wind rate is rela­ ditions. Knowledge of day-ahead hub-height wind is necessary to
tively high. Also, the integration of wind energy into the existing grids is develop wind farms, operational control, and grid scheduling.
still a big challenge due to the substantial variability of wind, especially Many prior studies have treated the estimation of wind using sta­
over complex terrain regions. tistical methods. For example, autoregressive, moving average, and
Most wind farms in Japan are scattered in mountainous areas where autoregressive integrated moving average models [3] are primarily
wind conditions are more complicated than those in other countries, employed. These methods extract information from available

* Corresponding author.
E-mail address: yuanliang@cuit.edu.cn (L. Yuan).

https://doi.org/10.1016/j.seta.2022.101995
Received 12 June 2021; Received in revised form 8 December 2021; Accepted 7 January 2022
Available online 19 January 2022
2213-1388/© 2022 Elsevier Ltd. All rights reserved.
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

observations, and thus the forecasting skill maintains a high accuracy As the objective of this study, a multi-scale model is built by coupling
only a few steps forward [4]. a mesoscale NWP model and a computational fluid dynamics (CFD)
Artificial intelligence (AI) techniques have become of particular in­ model to obtain hub-height wind forecasts for each turbine in the wind
terest during the last decade, specially those based on data-driven farm of interest, considering effects of local terrain and surface rough­
methods and universal approximators such as artificial neural net­ ness. The motivation of this coupling procedure is that implementing
works (ANN). These methods have the remarkable ability to extract either model (NWP or CFD) independently over complex terrain shows
patterns and detect trends from complicated or imprecise data. [5] and significant deficiencies. On one hand, the mesoscale model is not able to
[6] have employed techniques based on ANNs for wind speed predic­ adequately represent local terrain with significant variance, a feature
tion, while [7] included satellite-derived information in an ANN model that is vital to simulating wind flow close to the surface. On the other
to forecast atmospheric variables. Fuzzy logic models [8–10] and the hand, although CFD models can capture the local effects caused by
Jaya algorithm-based support vector machine model [11]have also been complex terrains, the lack of actual initial/boundary conditions makes
implemented. their application to real prediction ofwind impossible.
These data-driven methods are often used in a hybrid fashion with There already exists a few studies [23–27] dealing with combination
different techniques in order to improve their performance. [12] intro­ of NWP and CFD models, however, most of which are based on idealized
duced a forecasting model named ANN–MC involving ANN and Markov and controlled boundary conditions or low-resolution NWP output,
chain (MC) for the sake of obtaining accurate wind prediction in a few which cannot provide the turbine-specific day-ahead forecast, especially
seconds in the future. Mode decomposition techniques are also often where the weather condition varies significantly. Being a continuation
coupled with Long-Short term memory (LSTM) networks, which are a of our existing work [28] to establish a coupled prediction system of
type of ANN specially designed for time-series. [13] performed VMD- surface wind for wind farms in Japan, we have improved the coupling
SSA decomposition on demodulated wind speed time series and used procedure with more variables and comprehensively evaluated the
LSTM networks to predict every component, while [14] used a similar proposed multi-scale model for twelve turbines at a new wind farm
approach but with complete ensemble empirical mode decomposition (shown in Fig. 1) in Houhoku, Japan, during both summer and winter.
(CEEMD) as the decomposition step. In addition, CEEMD have also been As a result, turbine-specific forecasts can be derived. Validation results
coupled with stacking-ensemble learning approach and machine show that this coupled model aids to facilitate the day-ahead hub-height
learning models to develop a hybrid decomposition-ensemble learning wind forecast in Japan, a country with mostly rugged and mountainous
model [15] for multi-step ahead wind energy forecasting. terrain out of which approximately 66% is covered by forest.
It should be stressed that most of these approaches usually consider We consider the following points to be the main contributions of the
only historical wind speed series in one or several observation sensors, paper:
and therefore the atmospheric dynamics have not been taken into ac­
count, which is critical to the day-ahead prediction. • A multi-scale model is developed to predict day-ahead wind speed
Numerical weather prediction (NWP) models (e.g., Weather under complex terrain conditions. A very high resolution (i.e., 500
Research and Forecasting (WRF)), which are built based on physical m) NWP model and a microscale CFD model are coupled to solve
principles, are usually the optimal option for short-term wind prediction highly fluctuated wind flows.
[16,17]. However, the accuracy of near-surface wind forecasts derived • This work is the first effort to combine models with different scales
from NWP models remains relatively low due to the uncertainties and for the purpose of operational use for a wind farm located over
numerical errors, especially over complex terrain. This kind of situation complex terrains in Japan, to the best of the authors’ knowledge. We
has led to the combination of NWP models and statistical approaches sample two months of hourly forecasts from the proposed model to
such as theKalman filter algorithm [18], and statistical approaches that show its superior performance when compared with an NWP model.
act as a post-processing step [17,19–22]. Most of these studies employed Also, the reduced computational cost make this multi-scale model a
a low resolution NWP model (3∼ km), and the wind observations used to promising routine tool for providing day-ahead predictions.
calibrate the NWP forecasts were taken from only one or two meteoro­ • A comparison with statistical methods indicates that this coupling
logical towers. However, employing the similar methods to a wind farm procedure has a superior ability against most post-processing algo­
located over complex terrain areas is hard to obtain accurate forecast for rithms or AI techniques to reduce random uncertainties.
each individual turbine.

(a) (b)

Fig. 1. (a) Map showing the parent and two nested domains for the WRF model. The triangle stands for the wind farm of interest in Houhoku, Japan. (b) Elevation
(m) map of Houhoku wind farm showing the location of the 12 turbines.

2
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Next, in Section 2, the details of the dataset, the physical models, and study we employ RANS since the microscale air motion in the atmo­
the coupling procedure are described. Section 3 analyzes the perfor­ spheric boundary layer (ABL) is considered incompressible turbulent
mance of the proposed multi-scale model and a brief discussion is given. flow. The steady and incompressible RANS equations are,
The paper ends with concluding remarks in Section 4.
∂ui
= 0, (1)
∂xi
Model, Methodology and Data
[ ( )]
∂ui ∂p ∂ ∂ui ∂uj ∂ ′ ′
This section describes the mesoscale NWP model and microscale CFD ρuj = − + μ + − ρu u , (2)
∂xj ∂xi ∂xj ∂xj ∂xi ∂xj i j
model used to obtain the forecasts for the region of interest. Throughout
this paper, we refer to the direct output from the NWP model as either where u is the velocity, i and j indicate different directions, ρ is the air
WRF-alone or WRF500 forecast. This section also introduces the Global density, μ is the dynamic viscosity, p is the pressure, and (− ) denotes the
Forecasting System (GFS) data, and the observed data at hub height used Reynolds-averaged quantity. The last term in (2) represents the effect of
to build and validate the performance of the proposed model. sub-grid turbulent eddies, which can be calculated with the standard
k − ∊ turbulence model described in [39]. The turbulent viscosity is μt =
WRF model configuration 2
Cμ k∊ , where Cμ = 0.09, k is the turbulent kinetic energy, and ∊ denotes
the dispassion rate. In this study, all of these equations are solved
As a component of the multi-scale model, the meso-scale WRF model
numerically using the Open Source Field Operation and Manipulation
with Advanced Research WRF (ARW) version 4.0.1 is chosen in this
model (OpenFOAM, ver.7), which is one of the components of the pro­
study; a high-resolution regional model based on a fully compressible
posed multi-scale model. The second order upwind schemes for velocity,
and non–hydrostatic dynamic core. Using the outputs from global-scale
turbulence kinetic energy (TKE), and turbulence dissipation are chosen.
models (e.g., GFS and European Centre for Medium-Range Weather
More detailed information of OpenFOAM can be found in [40].
Forecasts Integrated Forecasting System) as initial/boundary condi­
tions, it can generate high-resolution forecasts in both spatial and tem­
poral spaces.
We set the forecasting domain following the recommendations in
[29,30]. As shown in Fig. 1a, one parent and two nested domains (d02
and d03) with a horizontal resolution of 6.0 km, 1.5 km, and 0.5 km,
respectively, are chosen in this paper; meanwhile, 35 stretched eta levels
in vertical for all domains are adopted. The location of the wind farm of
interest is marked with a red triangle in Fig. 1a, and the surrounding
topography is illustrated in Fig. 1b.
The main physical WRF options used in this work are summarized in
Table 1. These are the Asymmetric Convective Model version 2 (ACM2)
planetary boundary layer (PBL) scheme [31], the Noah land surface
model [32], the Eta similarity surface layer scheme [33], Dudhia
shortwave radiation [34], Rapid Radiative Transfer Model (RRTM)
longwave radiation [35], the new Kain–Fritsch convective parameteri­
zation [36], and WRF Single-Moment 6-class (WSM6) microphysics
parameterization scheme [37].
The WRF model needs to be re-initialized each day to output the day-
ahead forecasts. In this study, we run the WRF model on “cold-start”
mode at 18:00 UTC each day, and each run lasts thirty hours. The first
six hours are regarded as the spin-up period [38].Model output has a
constant interval of one hour.

Microscale model

Approaches such as direct numerical simulation (DNS), large eddy


simulation (LES), and Reynolds-averaged Navier–Stokes (RANS) are
usually used to solve the Navier-Stokes equations. DNS and LES require
solving the extensive range of temporal and spatial scales of a turbulent
flow; thus, a significant amount of computational resources are needed,
making them virtually impossible for practical applications. As a result,
more computationally affordable RANS modeling is the most common
and widespread approach in engineering applications. Therefore, in this

Table 1
Physics options used in WRF predictions for all three domains.
Physical schemes Option

PBL scheme ACM2


Land surface model Unified Noah Land Surface Model
Surface Layer Eta similarity
Shortwave Radiation Dudhia
Longwave Radiation RRTM
Cumulus convection Kain-Fritsch
Fig. 2. A schematic diagram of the multi-scale model for day-ahead wind
Microphysics WSM6
speed prediction..

3
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Table 2 speed, and TKE, have been included. To produce the OpenFOAM
Description of models adopted in this work. boundary conditions, we extracted information from WRF outputs with
Models Description a resolution of 500 m through employing the Cressman interpolation
algorithm, which is a method for progressive revisions to interpolate
WRF500 WRF raw forecasts with a high-resolution of 500 m.
WRF_OF100 Coupled model of which the horizontal resolution of OpenFOAM is discrete points (WRF grids in this study) into the regular grid ones within
100 m. small errors. It is generally simple, computationally fast and more ac­
WRF_OF050 Coupled model of which the horizontal resolution of OpenFOAM is curate than other interpolation methods such as bilinear interpolation.
50 m. The types of boundary conditions used in this work are summarized
in Table 3. The simulation reaches to the steady state when the residual
Multi-scale model: coupling WRF and OpenFOAM is less than 10− 4 . The choice of inlet and outlet patches depends on the
prevailing wind direction calculated using the output of the mesoscale
As mentioned earlier, the complex terrain effect plays a critical role component at the height of 500 m. We believe that this coupling pro­
in modifying wind speed and direction in a wind farm site. Under­ cedure results in more representative boundary conditions of local
standing the wind profile and turbulence is much more complicated than complex terrain than ideal wind profiles or information from coarser
in a flat area. Although NWP models can predict wind regimes at an mesoscale NWP models.
extensive range of scales from several kilometers to hundreds of meters, Fig. 3 displays the 20 km × 20 km domain of WRF_OF100. The
they cannot cope with the high spatial resolutions (e.g., < 500 m) domainoverlaps with a region of 40 × 40 grid points of the WRF model.
necessary to resolve small, yet often crucial, flow details occurring at a This kind of configuration is beneficial to extract sufficient information
wind farm over the complex terrain. CFD models have proven successful from the mesoscale model as initial/boundary conditions to drive the
in resolving wind flow features at a smaller scale. Nevertheless, the use OpenFOAM model. Vertically, 30 stretched levels with an expansion
of idealized or synthetic initial/boundary conditions for a CFD model ratio of 1.94 are chosen. As a result, the mesh close to the surface is
leads to unrealistic flow, which cannot be used directly for the purpose highly refined (shown in Fig. 4) to accurately capture the local details
of operational prediction. Hence, coupling NWP and CFD models can caused by topographic features. The total number of grid points is 1.2
result in a better modeling approach to predict day-ahead wind for a and 4.8 million for WRF_OF100 and WRF_OF050, respectively.
wind farm located in mountainous areas. The basic idea is shown in
Fig. 2. Data
This study proposes using the mesoscale WRF predictions as the
initial guess and boundary conditions for the OpenFOAM model. As Three different types of data were used: background data for driving
shown in Table 2, two kinds of resolutions (i.e., 100 and 50 m) for the the WRF model, terrain data to generate the grids for the multi-scale
micro-component model are chosen. Compared to our previous study model, and observed data for verifying the forecasting skills of the
[28] more variables, i.e., horizontal wind speed, direction, vertical proposed model.
Two months’ (January and July of 2019) National Centres for

Table 3
The boundary condition types used for micro-scale component of the proposed multi-scale model.
Boundary Velocity Pressure TKE Epsilon

inlet_patch fixedValue zeroGradient fixedValue fixedValue


outlet_patch inletOutlet fixedValue zeroGradient zeroGradient
ground_patch fixedValue zeroGradient kqRWallFunction epsilonWallFunction

Fig. 3. Calculation domain of the OpenFOAM model. Turbines 1–7 are located around location A and 8–12 are located at B..

4
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Fig. 4. An enlarged part of the OpenFOAM domain showing the horizontal and vertical mesh structures..

Environmental Prediction GFS forecasting data are used. The time in­ VMD was introduced by [44] as a superior alternative to the well-
terval of this dataset is three hours, and the horizontal resolution is known Empirical Mode Decomposition (EMD). In VMD decomposition
0.25 × 0.25 degree (approximately 27 km) for all variables. all mode waveforms and their corresponding central frequencies are
The WRF model uses Global Multi-resolution Terrain Elevation Data calculated simultaneously, with each resulting mode having limited
2010 dataset (30-arc-second spatial resolution), which contains eleva­ bandwidth and specific sparsity properties. VMD is formulated as a
tion data for the globe collected from various sources. Besides, the constrained variational problem. The modes are calculated by using the
Shuttle Radar Topography Mission (SRTM) data set, which has a reso­ Hilbert transform and the width of the sub-modal signal are estimated
lution of 3-arc-second, is chosen to generate the structured mesh cells for from the demodulated signal. Each mode and frequency are obtained
both WRF_OF100 and WRF_OF050. SRTM data is in the WGS84 datum; such that they minimize the variational problem.
therefore, the coordinates should be transformed from the Cartesian As for the time series prediction method, LSTM [45] is one type of
system to Universal Transverse Mercator (UTM). recurrent neural network (RNN), targeted at the prediction of time series
To verify the performance of the model predictions we use quality- and other sequential data. The main difference with a common artificial
controlled nacelle wind observations from twelve turbines of the wind neural network is that RNNs are able to share information across
farm of interest (red triangle in Fig. 1), located in the Houhoku area. different prediction steps by using feedback connections. In the case of
These data sets are available from 2015/01/01 00:00 (JST) to 2019/12/ LSTMs, three gates named “input”, “output” and “forget” control the
31 23:50 (JST), with an interval of one hour. information that must be shared with future predictions, as well as the
information that must be forgotten. This makes LSTMs very useful in
VMD-LSTM keeping track of patterns in the data even if these happen to occur only
sporadically.
One of the chosen validation methods is a combination of mode
decomposition and neural network-based time series prediction. His­ Kalman filter algorithm
torical time series of wind speed are decomposed into its corresponding
modes, while the time series prediction method is used on each mode to Kalman filter algorithm is also employed in this study to further
forecast its next time step. The final prediction is obtained by recom­ validate the performance of the proposed multi-scale model. It has been
posing all predicted modes. In this paper we employ VMD and LSTM widely used in various fields because it can operate recursively on
networks respectively. This combination has been employed in other available data streams to produce a statistically optimal estimation of
studies regarding hybrid methods, both for wind speed [13,41] and the underlying system state, requiring only a small computational cost.
wind power [42,43]. As most machine learning-based models, these Similar to ([46,19]), the main mathematical equations which are used to
implementations are based solely on historical data and do not comprise estimate and remove the errors in the wind speed forecasts, are
any physical model. described as follows.

5
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

( )
x t+δt|t = ̂
̂ x t|t− δt + Gt yt − ̂
x t|t− δt , (3) Effects of complex terrain on hub-height wind forecasts

pt− + σ 2η,t The input of terrestrial data is critical for limited-area mesoscale
(4)
δt
Gt = 2
, models (e.g., WRF) to accurately produce predictions. The wind vector
δt + σ η,t + σ ∊,t
pt− 2
at 00:00 UTC 01 January 2019, and the corresponding terrain conditions
( )( ) for WRF500, WRF_OF100, and WRF_OF050 are plotted in Fig. 6, in order
pt = pt− δt + σ 2η,t 1 − Gt , (5) to illustrate the struggle of the WRF model to detect differences within
the wind farm over complex terrain. Clearly, WRF500 (Fig. 6a) considers
where xt and yt are the forecasting bias (unknown) and forecasting error the region around the wind farm as a much smoother surface compared
(forecast minus observation) at time t, δt is the time lag, Gt and pt denote to the WRF_OF100 and WRF_OF050 (Figs. 6b and 6c), both which show
the Kalman gain and expected mean square error, and ηt and ∊t are white largely mountainous topography. Correspondingly, the wind vector
noise with variances σ 2η,t and σ2∊,t , respectively. Given a initial guess of p0 among the three models is different.
and G0 , this Kalman filter will quickly converge. The detailed informa­ Another possible explanation for the significant difference in terrain
tion can be found in [46]. between WRF500 and WRF_OF100 or WRF_OF050 is the mismatch of
coordinate systems between the input of terrestrial data for WRF500 and
Results and Discussion the WRF model itself. [48] has pointed out that this distinction is often
neglected during WRF preprocessing (remapping to WRF grids); how­
First, this section introduces the wind conditions in the site of in­ ever, it can cause geolocation shift that can exceed 20 km at mid-
terest. Then we analyze the issues with using only the WRF model to latitudes.
predict wind speed at hub height over complex terrain areas. Next, the To further illustrate the local effects of complex terrain, the wind
output of the multi-scale model is verified during both summer (July) speed among twelve turbine sites for four cases, i.e., 00:00, 06:00, 12:00,
and winter (January). Finally a short discussion is given. 18:00 UTC 01 January 2019, is investigated. Fig. 7 displays the box plots
of wind speed for the three models. It is observed that WRF raw forecasts
(WRF500) largely overestimate the wind speed compared to the corre­
Wind conditions in the site of interest sponding observations for all four cases, whilst the outputs from
WRF_OF100 and WRF_OF050 are much closer to observations. The big
The Houhoku wind farm is situated in the western part of Japan and interquartile range (IQR) for observed values indicates that the degree of
consists of twelve wind turbines (hub height is 85 m). We have per­ dispersion is apparent, reflecting that the role of local terrain cannot be
formed a statistical analysis based on a full year (2019) of actual data neglected. However, the local effects of complex terrain in WRF500 do
described in Section 2.4. The two-parameter Weibull distribution [47] is not emerge as expected, especially for the case at 06:00 UTC, because
chosen to characterize wind regimes because it has been found to pro­ the IQR is much smaller than that for observed wind. In contrast, the
vide a good fit with measured wind data. The probability density results of WRF_OF100 and WRF_OF050 are more similar to observations,
function f is given by and the mean wind speed is also close to the observed value. Note that
( ) ( ( ) ) WRF_OF050 outperforms the other two models.
k (w)k− 1 w k
f w = exp − , (6)
c c c
Validation of the multi-scale model
where w, k, and c are the wind speed, Weibull shape factor, and scaling
velocity, respectively. From Fig. 5, we can easily find that the shape Firstly, in Fig. 8 we present four cases of WRF-alone and the multi-
factor is 1.92, which translates into relatively fluctuating winds due to scale model wind speed time series, together with observed values,
complex terrain conditions. Also, the data analysis reveals that the mean during winter. Each case lasts a period of ten days from 00:00 UTC 01 to
wind speed and direction is 5.37 m/s and 235 degrees, respectively, 23:00 UTC 10 January 2019. These cases are selected to highlight the
during the sampled period. The largest mean wind speed (6 m/s) occurs advantages of the multi-scale compared to the WRF-alone forecasts.
from the south-southwest. Although there are occasional significant errors, we find that the

Fig. 5. Weibull distribution of wind speed (left) and wind rose (right) at the wind farm of interest. k, c, SpdAve, and DirAve represent the Weibull shape factor, scale
factor, the mean wind speed, and mean direction, respectively. The bar in the left panel shows the frequency of the observed wind speeds. The value at the end of
each directional line (right panel) is the mean wind speed at that direction.

6
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Fig. 6. Wind vector of WRF500 (a), WRF_OF100 (b), and WRF_OF050 (c) at 00:00 UTC 01 January 2019. Shaded contours indicate the terrain height for different
models. Triangles stand for the wind turbines of Houhoku wind farm.

Fig. 7. Box plots of wind speed for three different models, as well as the corresponding observations, at 00:00, 06:00, 12:00, and 18:00 UTC 01 January 2019. The
purple triangle represents the 12-turbine mean of the observed wind speed.

WRF500 wind speed coincides nicely with the observation through all overestimation of WRF-alone prediction and can bring the values closer
four cases in general, indicating that the WRF500 can predict the local to the observations for all four cases, surpassing WRF_OF100. It worth
wind speed in the wind farm of interest during winter. Coupling with the noting that the variation of WRF500 wind forecasts among four turbines
microscale OpenFOAM model leads to the improvement of the fore­ is much smaller than WRF_OF050.
casting skills of WRF500. It can be easily observed that the impact of The scenario in summer is depicted in Fig. 9. The ten-day (from 00:00
WRF_OF100 on the WRF-alone forecasts seems minor at sites No.2 UTC 01 to 23:00 UTC 10 July 2019) forecasts derived from WRF500,
(Fig. 8a) and No.5 (Fig. 8b) while it is significant for sites No.9 (Fig. 8c) WRF_OF100, and WRF_OF050 for all twelve turbines are analyzed.
and No.12 (Fig. 8d). Moreover, WRF_OF050 successfully corrects the Again, the WRF500 model demonstrates its capabilities for predicting

7
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Fig. 8. Hourly WRF-alone forecasts (black) and corresponding observations (red) of wind speed at hub height of No.2 (a), No.5 (b), No.9 (c), and No.12 (d) turbines
for the 10-day period from 00:00 UTC 01 January to 23:00 UTC 10 January 2019. The blue and green lines present the predictions derived from WRF_OF100 and
WRF_OF050, respectively.

Fig. 9. Hourly WRF-alone forecasts (black) and corresponding observations (red) of wind speed at hub height of twelve turbines for the 10-day period from 00:00
UTC 01 July to 23:00 UTC 10 July 2019. The blue and green lines present the predictions derived from WRF_OF100 and WRF_OF050, respectively. (a)-(l) shows the
turbines No.1-No.12.

the local wind speed in summer. Consistent with the situation in winter, models as well as the WRF500, three statistical parameters, i.e., mean
the WRF500 forecast for the twelve turbines has very similar patterns, error (ME), mean absolute error (MAE), and root mean square error
which might suggest that the local terrain effect has been neglected to (RMSE) are employed. These are given by
some extent due to the coarse resolution. When comparing with the

Nt
results of the two multi-scale models, we see the remarkable advantage ME =
(
pt − at
)/
Nt , (7)
of coupling mesoscale model with microscale model in reducing the t=1
forecasting errors, especially the systematical error. ⃒ ⃒
Overall, WRF_OF050 shows the best performance, both in winter and ∑Nt ⃒

⃒/

summer. On the other hand, WRF_OF100 offers moderate improvement MAE = ⃒pt − at ⃒ Nt , (8)
t=1
⃒ ⃒
and sits mostly between WRF500 and WRF_OF050; although in the
summer case, it can get closer to the observations during days 4–5. All (

Nt
)
/
these results show the proposed multi-scale model outperforming the RMSE = sqrt (pt − at )2 Nt , (9)
WRF-alone model, and it can be seen that the higher resolution of the t=1

microscale component we adopt, the larger improvement can be


obtained. where t is the time point and Nt is the total number of time points during
Next, to quantitatively assess the performance of the multi-scale the verifying periods. p and a represent the prediction and actual value,

8
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

respectively. For both summer and winter cases, a one-month length of For the winter scenario, it can be seen that the ME of WRF-alone pre­
data is used to compute those statistical parameters. The results are diction varies between the 12 turbines (Fig. 11a), from 1.87 m/s to 3.43
summarized in Figs. 10 and 11, and Tables 4 and 5. m/s. Both WRF_OF100 and WRF_OF050 can reduce the overestimation
Fig. 10 shows the diurnal cycle of 12-turbine averaged wind speed of WRF-alone for all turbines, suggesting that the multi-scale models
and MAE values for WRF500, WRF_OF100, and WRF_OF050. It is have a solid ability to alleviate the systematic error of the WRF model.
observed that the diurnal cycle of the observed wind is not evident for The largest reduction can be found at turbine No.8 with 99.98%, given
both winter (Fig. 10 and summer (Fig. 10c) cases, where the wind at by the WRF_OF050 model. The12-turbine mean is 93.08%. Consistent
local night-time (09:00–21:00 UTC) is slightly larger than that at day- with ME, results for MAE and RMSE (displayed in Figs. 11b and 11c)
time in winter, while the phenomenon in summer is opposite. Again, further demonstrate that multi-scale models are superior to WRF500 for
the overestimation of WRF500 can be observed during both winter and day-ahead hub height wind prediction. Note that WRF_OF050 out­
summer, and the multi-scale models can primarily reduce it, in partic­ performs WRF_OF100 for all turbines.
ular, WRF_OF050. The multi-scale model (i.e., WRF_OF050) forecasts’ Similarly, in summer, the systematic error of WRF500 has been
variability in space is supported by the behaviour of box plots in reduced by WRF_OF100 and WRF_OF050 as well (Fig. 11d), with the 12-
Figs. 10a and 10c, and it is more significant in winter than the case in turbine-mean of relative improvements being 47.39% and 83.44%,
summer. It also should be mentioned that the forecasting skills of the respectively, although such improvement is slightly smaller compared to
three models do not degrade too much with the increase of forecasting those in winter. The MAE and RMSE displayed in Figs. 11e and 11f also
lead time, indicating the advantage of physical-based models against the demonstrate that coupling a mesoscale NWP model with a microscale
pure statistical models for day-ahead wind prediction. CFD model has the capability of making up for the drawbacks of
The MAE is exhibited in Figs. 10b and 10d for winter and summer, employing the mesoscale model solely over the complex terrain area.
respectively, to show the quality of the forecasts. Again, the WRF-alone Clearly, WRF_OF050 is superior to WRF_OF050 in regard to the reduc­
model shows the largest values of MAE, while WRF_OF050 shows the tion of both systematic errors and random uncertainties.
smallest, at all times. There is a clear trend that the WRF500 MAE values From the above discussions, we may conclude that employing the
at local night-time are larger than local day-time. The increase of MAE multi-scale models proposed in this study can substantially improve the
starts around 18:00 JST (09:00 UTC), which might reveal that the WRF accuracy of the WRF-alone model in forecasting the hub-height wind
model’s weakness of precisely capture the physical processes during the field in the target wind farm site of interest, where the terrain condition
evening transition [49] over complex terrain region. However, the is complex. The more satisfactory resolution of the micro-scale compo­
multi-scale models also cannot effectively improve such a situation. nent we have, the better improvement against the WRF-alone model can
More adaptive local parameterization schemes may be needed. be obtained.
We show the performance of WRF500, WRF_OF100 and WRF_OF050
at each turbine site in Fig. 11. Also, to quantitatively prove the advan­
tages of the multi-scale models, we have compiled in Table 4 and Table 5
the summarized results for the cases of winter and summer, respectively.

Fig. 10. (a) and (c) show the diurnal cycle of 12-turbine averaged wind speed of WRF500, WRF_OF100, WRF_OF050 and the corresponding actual values during
winter and summer, respectively; box plots signify the variation of WRF_OF050 wind speed between twelve turbines. (b) and (d) are the the diurnal cycle of wind
speed MAE for different models in winter and summer, respectively.

9
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Fig. 11. A comparison of the ME, MAE, and RMSE for wind speed at hub-height of the WRF_OF100 (solid light blue bar) and WRF_OF050 (solid green bar) pre­
dictions with respect to the raw WRF-alone prediction (solid white bar) for the twelve turbines of the Houhoku wind farm, during winter (left panels) and summer
(right panels). The averaged results of three statistical parameters are also shown.

Table 4
Improvement of ME, MAE, and RMSE (m/s) for twelve turbine sites at the Houhoku wind farm on January 2019.
ME MAE RMSE

Site WRF500 WRF_OF100 WRF_OF050 WRF500 WRF_OF100 WRF_OF050 WRF500 WRF_OF100 WRF_OF050
Improv. (%) Improv. (%) Improv. (%) Improv. (%) Improv. (%) Improv. (%)

No.1 2.30 28.08 89.16 2.65 5.35 27.69 3.26 3.11 25.22
No.2 2.34 15.95 96.50 2.69 -2.16 34.40 3.25 -3.93 30.94
No.3 2.08 3.02 90.83 2.51 -6.69 30.01 3.06 -8.52 26.32
No.4 1.90 2.87 85.02 2.30 -7.56 23.48 2.81 -10.15 19.11
No.5 2.48 26.59 99.72 2.82 9.92 41.35 3.38 7.75 36.87
No.6 2.87 29.48 94.41 3.05 20.00 42.80 3.59 15.18 37.71
No.7 3.02 30.04 90.06 3.25 15.29 46.55 3.92 12.80 42.85
No.8 2.85 47.02 99.98 3.02 28.06 48.99 3.53 24.35 44.43
No.9 3.23 50.89 94.88 3.36 36.25 57.26 3.89 31.84 52.82
No.10 2.38 40.33 91.70 2.68 19.06 43.27 3.16 15.72 37.78
No.11 2.19 28.60 96.10 2.53 10.21 38.43 3.04 6.56 33.52
No.12 3.11 39.10 88.63 3.25 25.88 53.38 3.80 21.95 49.63
Mean 2.56 28.50 93.08 2.84 12.80 40.64 3.39 9.72 36.43

Comparison with statistical methods chitecture: one input layer of size 120, two hidden LSTM layers with 64
and 128 nodes respectively, and one output layer of size 1. The networks
VMD-LSTM are train to predict only one time step in the future. The 24-h forecast is
This subsection will briefly describe the implementation details of obtained in a recursive manner that used in [13]; after each prediction
the VMD-LSTM validation method. In order to obtain a target 24-h the first entry in the input vector is discarded, the predicted value is
forecast, the wind data of the 5 previous days (120 values) was taken appended and the resulting input vector is fed to the LSTM to obtain the
as the main feature for training. Feature and target vectors are decom­ next prediction. After all LSTMs have produced the 24-h prediction for
posed using the VMD algorithm and a k = 5 number of modes is their corresponding components, the results are combined and a
obtained.k +1 LSTM networks are trained independently to learn to reconstructed wind speed forecast is obtained. g Out of five years of
predict each mode, plus the residue. All networks have the same ar­ observed wind speed data, four were used for training and one for

10
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Table 5
Improvement of ME, MAE, and RMSE (m/s) for twelve turbine sites at the Houhoku wind farm on July 2019.
ME MAE RMSE

Site WRF500 WRF_OF100 WRF_OF050 WRF500 WRF_OF100 WRF_OF050 WRF500 WRF_OF100 WRF_OF050
Improv. (%) Improv. (%) Improv. (%) Improv. (%) Improv. (%) Improv. (%)

No.1 1.96 61.00 74.78 2.38 19.51 32.07 2.89 16.46 19.16
No.2 1.84 58.32 76.83 2.28 14.96 26.62 2.74 11.13 12.01
No.3 1.50 37.13 70.61 2.16 10.24 16.32 2.72 8.21 5.63
No.4 1.60 28.73 76.33 2.45 6.63 28.73 3.03 5.10 17.69
No.5 1.66 53.90 74.91 2.14 8.06 24.15 2.72 7.07 16.14
No.6 1.65 47.81 78.08 2.08 7.54 27.05 2.65 7.67 18.42
No.7 2.11 30.43 95.17 2.36 11.69 36.55 2.99 9.81 29.07
No.8 2.49 50.47 88.52 2.71 18.76 39.15 3.41 13.60 36.54
No.9 2.60 55.46 87.59 2.79 26.93 44.25 3.69 25.00 44.17
No.10 1.98 51.12 99.48 2.28 9.59 28.62 2.96 6.46 25.69
No.11 1.58 46.27 95.54 2.06 10.41 17.55 2.65 4.99 14.29
No.12 2.52 48.00 83.44 2.72 25.91 41.91 3.52 22.27 40.44
Mean 1.96 47.39 83.44 2.37 14.19 30.25 3.00 11.48 23.27

Fig. 12. Profiles of the diurnal cycle of 12-turbine mean ME (a) and RMSE (b). Dotted lines stand for results of VMD-LSTM during winter (blue) and summer (red),
while solid lines represent results of WRF_OF050.

validation. The networks were trained using MSE as a loss function and process the WRF-alone forecasts.
the Adapted Momentum Estimation method as an optimizer [50]. Each We implement the Kalman filter to the forecasts of the WRF-alone
component was normalized independently before training. model and WRF_OF050 at each wind turbine separately, during both
Fig. 12 shows the comparison of diurnal cycle of 12-turbine mean of winter and summer. The first 21 days are chosen as a training period,
ME and RMSE between VMD-LSTM and WRF_OF050, during both winter and the left ten days for validation. Although the Kalman filter pro­
and summer. It is observed that both methods have much smaller sys­ cedure can be applied independently to every prediction lead time, we
tematic errors (Fig. 12a) compared to the WRF-alone forecasts (shown in adopt one step (1 h) forward in this study, considering the length of the
Figs. 10a and 10c). For the winter scenario, 24-h averaged MEs are data sample. To evaluate the performance of Kalman filtered forecasts
0.004 m/s and 0.092 m/s for WRF_OF050 and VMD-LSTM, respectively, for alleviating random error explicitly, the centred RMSE (CRMSE),
while they are − 0.120 m/s and 0.036 m/s during summer. These results which is the random component of the RMSE [51],
indicate that VMD-LSTM shows roughly equivalent skill to WRF_OF050
CRMSE2 = RMSE2 − ME2 , (10)
in terms of the systematical errors. Nevertheless, RMSE values (Fig. 12b)
indicate that physically-based WRF_OF050 shows a better capability
is used. The 12-turbine averaged results are summerized in Fig. 13,
against VMD-LSTM. The forecasting skill of VMD-LSTM decreases fast
showing the WRF-alone (diamond) and the multi-scale WRF_OF050
with forecasting lead time, especially after 12 h, during both winter and
forecasts (asterisk), and the post-processing forecasts by combining
summer. On the contrary, the RMSE of WRF_OF050 maintains a nearly
Kalman filter, i.e., “WRF500 + KF” and “WRF_OF050 + KF”. The dotted
consistent level throughout.
grey circular line has a radial distance from the origin (0, 0) that equals
the RMSE value. Consistent with the conclusion in Section 3.3, WRF500
Kalman filter
shows relatively large RMSEs being 3.29 m/s and 1.72 m/s in winter and
As mentioned earlier, many statistical methods, such as the Model
summer respectively, and it can be reduced by employing WRF_OF050.
Output Statistics method and Kalman filter, are usually employed as
A closer look at Fig. 13a let us find that the contribution of ME and
post-processing tools to alleviate the forecasting errors of NWP models
CRMSE to RMSE of WRF500 is much the same. Combining WRF and
in the wind energy field. These methods can combine available historical
OpenFOAM leads to a considerable reduction of both systematical error
observations to bring the predicted values closer to the observations for
and random error. Specifically, for total reduction of RMSE by
each turbine in the wind farm, considering the local effects as we did in
WRF_OF050, the percentages are 79% for systematical error and 21% for
this work. To determine whether the coupling procedure proposed in
random error. Implementing Kalman filter to WRF-alone forecast (5-
this work is necessary, we also adopt the Kalman filter algorithm to post-
point star in Fig. 13a) also exhibits the improvement of RMSE; it even

11
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

Fig. 13. ME as a function of CRMSE for the cases of winter (a) and summer (b), showing the WRF-alone (diamond) and the multi-scale WRF_OF050 forecasts
(asterisk), and post-processing forecasts of Kalman filter.

performs better than WRF_OF050. However, that reduction mainly owes head WRF500, WRF_OF100, and WRF_OF050 forecasts. A doubling
to the improvement of systematical error, which corresponds to 89% in in calculation time is observed when increasing the microscale
the entire reduction of RMSE. That means that the Kalman filter plays a component’s resolution from 100 m to 50 m. Note that increasing the
crucial role in reducing the difference of central location of the forecast number of cores used for the microscale model will reduce the total
and the observation while minor impact to random uncertainties. wall clock time, enabling this multi-scale model (WRF_OF050) to be
Finally, it should be mentioned that “WRF_OF050 + KF” outperforms a practical tool.
others on reducing both ME and CRMSE, indicating that combining the • Most studies on simulating wind flow over complex terrains choose
multi-scale and Kalman filters has advantages against the direct imple­ the steady neutral conditions within the atmospheric boundary layer
mentation of Kalman filter to WRF-alone forecast. Similar results (ABL). In this work, non-neutral conditions are adopted because the
observed in Fig. 13b further support the above conclusion. boundary conditions come from WRF real-time predictions. How­
ever, it is essential to note that some points in the boundary of the
micro-scale model are out of the WRF lowest level due to the dif­
Discussion ference in the terrain conditions. Therefore, there is no enough in­
formation in those points when extracting from the WRF model. We
The proposed multi-scale models behave better than the mesoscale solve this drawback using the log wind profile under the neutral
WRF model from the validation above. However, on this multi-scale condition in this work.
model, some aspects need further discussion. • For operational use, the actual wind direction from the WRF model is
used to determine the patch types (e.g., inlet and outlet) of micro­
• The comparison of WRF_OF100 and WRF_OF050 indicates that a scale component of the multi-scale model.
better resolution of the microscale component for the multi-scale • This multi-scale model can be a promising candidate for imple­
model generates better forecasts at individual turbines site over a menting statistical post-processing approaches, most of which can
mountainous region. Following this principle, a higher resolution, only handle the systematical bias.
for example, 25 m, may lead to a better forecast compared to
WRF_OF050. However, as mentioned in Introduction section, we Conclusion
hope this model can be a helpful tool for operational use. Increasing
the resolution makes it difficult to afford the computational burden. The multi-scale model combining the advantages of a mesoscale
There is a comprise between accuracy and computational cost. NWP model and a microscale CFD model was introduced, to generate
Table 6 shows the wall clock time (hours) needed to finish the day- day-ahead forecasts with high resolution over the regions of complex
geographical features where the wind patterns are more fluctuating than
Table 6 usual.
The wall clock time (hours) needed to finish the day-ahead forecasts. The wall The proposed model has been validated in winter and summer using
clock time for the mesoscale WRF model is measured in a server with the 2.00 data at a wind farm of interest in Japan, where the terrain condition is
GHz Intel(R) Xeon(R) Platinum 8164 processor, while for micro-scale compo­ complex. It is observed that the multi-scale models consider the local
nent it is measured in a PC with the 3.00 GHz Intel i9-9980XE processor.
features and can substantially improve the accuracy of the WRF-alone
Models No. of cores for No. of cores for Total wall clock model in forecasting the hub-height wind field in winter and summer.
mesoscale component microscale component time (hours)
Specifically, compared to the WRF-alone forecasts, the 12-turbine
WRF500 50 — ∼1.3 averaged improvement of WRF_OF050 in ME, MAE, and RMSE is
WRF_OF100 50 12 ∼1.3 + ∼2.4 93.08%, 40.64% and 36.43%, respectively, in the winter case. Similarly,
WRF_OF050 50 12 ∼1.3 + ∼4.8 in summer, the improvement is 83.44%, 30.25%, and 23.27% in ME,

12
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

MAE, and RMSE, respectively. [13] Rodrigues Moreno S, Gomes da Silva R, Cocco Mariani V, dos Santos Coelho L.
Multi-step wind speed forecasting based on hybrid multi-stage decomposition
Also, we find that the higher resolution of the microscale component
model and long short-term memory neural network. Energy Conversion and
leads to a superior improvement against the WRF-alone model, although Management 2020;213:112869.
there is a comprise between accuracy and computational burden when it [14] Liu Z, Hara R, Kita H. Hybrid forecasting system based on data area division and
is expected to be a practical tool. deep learning neural network for short-term wind speed forecasting. Energy
Conversion and Management 2021;238:114136.
In addition, the results of VMD-LSTM and Kalman filter indicate that [15] da Silva RG, Ribeiro MHDM, Moreno SR, Mariani VC, dos Santos Coelho L. A novel
this coupling procedure has a competitive ability for day-ahead wind decomposition-ensemble learning framework for multi-step ahead wind energy
forecastingagainst most of AI-based methods or post-processing algo­ forecasting. Energy 2021;216:119174.
[16] Carvalho D, Rocha A, Gómez-Gesteira M, Santos CS. WRF wind simulation and
rithms to reduce forecasting errors, especially for random uncertainties. wind energy production estimates forced by different reanalyses: Comparison with
Finally, it is essential to note that across-scale wind forecasting is very observed data for Portugal. Applied Energy 2014;117:116–26.
challenging. The coupled model in this work was developed based on [17] Che Y, Peng X, Delle Monache L, Kawaguchi T, Xiao F. A wind power forecasting
system based on the weather research and forecasting model and kalman filtering
some simplifications, including the absence of Coriolis, buoyancy and over a wind-farm in japan. Journal of Renewable and Sustainable Energy 2016;8:
gravity forces, and a precise canopy model. In addition, a turbine model 013302.
may be needed to solve wake effects among different turbines. [18] Kalman RE. A new approach to linear filtering and prediction problems. Journal of
basic Engineering 1960;82:35–45.
[19] Monache LD, Nipen T, Liu Y, Roux G, Stull R. Kalman Filter and Analog Schemes to
CRediT authorship contribution statement Postprocess Numerical Weather Predictions. Monthly Weather Review 2011;139:
3554–70. https://doi.org/10.1175/2011MWR3653.1.
[20] Gneiting T, Raftery AE, Westveld AH, Goldman T. Calibrated Probabilistic
Yuzhang Che: Conceptualization, Methodology, Investigation, Forecasting Using Ensemble Model Output Statistics and Minimum CRPS
Writing - original draft, Writing - review & editing, Visualization. Estimation. Monthly Weather Review 2005;133:1098–118. https://doi.org/
Andrés A. Salazar: Writing - review & editing, Methodology. Siyue 10.1175/MWR2904.1.
[21] Bremnes JB. Probabilistic Forecasts of Precipitation in Terms of Quantiles using
Peng: Software, Visualization, Data curation. Jiafeng Zheng: Software, NWP Model Output. Monthly Weather Review 2004;132:338–47.
Formal analysis, Data curation. Yangruixue Chen: Writing - review & [22] Wilks DS. Statistical Methods in the Atmospheric Sciences. 3rd ed. Academic Press;
editing, Visualization. Liang Yuan: Conceptualization, Methodology, 2011.
[23] Zajaczkowski FJ, Haupt SE, Schmehl KJ. A preliminary study of assimilating
Formal analysis.
numerical weather prediction data into computational fluid dynamics models for
wind prediction. Journal of Wind Engineering and Industrial Aerodynamics 2011;
99:320–9.
Declaration of Competing Interest [24] Miao Y, Liu S, Chen B, Zhang B, Wang S, Li S. Simulating urban flow and dispersion
in beijing by coupling a cfd model with the wrf model. Advances in atmospheric
The authors declare that they have no known competing financial sciences 2013;30:1663–78.
[25] J. O’Sullivan, Modelling wind flow over complex terrain, Ph.D. thesis,
interests or personal relationships that could have appeared to influence ResearchSpace@ Auckland, 2012.
the work reported in this paper. [26] Blocken B, van der Hout A, Dekker J, Weiler O. CFD simulation of wind flow over
natural complex terrain: case study with validation by field measurements for Ria
de Ferrol, Galicia, Spain. Journal of Wind Engineering and Industrial
Acknowledgements Aerodynamics 2015;147:43–57.
[27] Temel O, Bricteux L, van Beeck J. Coupled wrf-openfoam study of wind flow over
We acknowledge the support of Clean Energy Factory Co., Ltd., complex terrain. Journal of Wind Engineering and Industrial Aerodynamics 2018;
174:152–69.
Tokyo, Japan, for providing the observation data. This work is sup­ [28] Jin J, Che Y, Zheng J, Xiao F. Uncertainty quantification of a coupled model for
ported by National Natural Science Foundation of China (grants No. wind prediction at a wind farm in japan. Energies 2019;12:1505.
41905084) and National Undergraduate Training Program for Innova­ [29] Warner TT, Peterson RA, Treadon RE. A Tutorial on Lateral Boundary Conditions
as a Basic and Potentially Serious Limitation to Regional Numerical Weather
tion and Entrepreneurship (grants No. 202010621003).
Prediction. Bulletin of the American Meteorological Society 1997;78:2599–617.
https://doi.org/10.1175/1520-0477(1997)078<2599:ATOLBC>2.0.CO;2.
References [30] Warner TT. Quality Assurance in Atmospheric Modeling. Bulletin of the American
Meteorological Society 2011;92:1601–10. https://doi.org/10.1175/BAMS-D-11-
00054.1.
[1] J. Lee, F. Zhao, Gwec global wind report 2019, Wind Global Energy Council, Tech.
[31] Pleim JE. A Combined Local and Nonlocal Closure Model for the Atmospheric
Rep. (2020).
Boundary Layer. Part I: Model Description and Testing, Journal of Applied
[2] Institute for Sustainable Energy Policies, Share of renewable energy electricity in
Meteorology and Climatology 2007;46:1383–95. https://doi.org/10.1175/
japan, 2019 (preliminary report), url:https://www.isep.or.jp/en/879/, 2020.
JAM2539.1.
[3] Torres JL, Garcia A, De Blas M, De Francisco A. Forecast of hourly average wind
[32] Chen F, Dudhia J. Coupling an Advanced Land Surface – Hydrology Model with the
speed with arma models in navarre [spain]. Solar energy 2005;79:65–77.
Penn State – NCAR MM5 Modeling System. Part I: Model Implementation and
[4] Hering AS, Genton MG. Powering up with space-time wind forecasting. Journal of
Sensitivity. Monthly Weather Review 2001:569–85.
the American Statistical Association 2010;105:92–104.
[33] Janić ZI. Nonsingular implementation of the mellor-yamada level 2.5 scheme in the
[5] X. Wang, G. Sideratos, N. Hatziargyriou, L.H. Tsoukalas, Wind speed forecasting for
ncep meso model. NCEP Office Note 2001;437:61.
power system operational planning, in: 2004 International conference on
[34] Dudhia J. Numerical Study of Convection Observed during the Winter Monsoon
probabilistic methods applied to power systems, IEEE, 2004, pp. 470–474.
Experiment Using a Mesoscale Two-Dimensional Model. Journal of Atmospheric
[6] Salcedo-Sanz S, Perez-Bellido AM, Ortiz-García EG, Portilla-Figueras A, Prieto L,
Sciences 1989;46:3077–107.
Correoso F. Accurate short-term wind speed prediction by exploiting diversity in
[35] Mlawer EJ, Taubman J, Brown PD, Iacono MJ, Clough SA. Radiative transfer for
input data using banks of artificial neural networks. Neurocomputing 2009;72:
inhomogeneous atmospheres: RRTM, a validated correlated-k model for the
1336–41.
longwave. Journal of Geophysical Research 1997;102.
[7] Deo RC, Şahin M. Forecasting long-term global solar radiation with an ann
[36] Kain JS. The Kain-Fritsch Convective Parameterization: An Update. Journal of
algorithm coupled with satellite-derived (modis) land surface temperature (lst) for
Applied Meteorology 2004;43:170–81.
regional locations in queensland. Renewable and Sustainable Energy Reviews
[37] Hong S-Y, Noh Y, Dudhia J. A New Vertical Diffusion Package with an Explicit
2017;72:828–48.
Treatment of Entrainment Processes. Monthly Weather Review 2006;134:2318–41.
[8] Potter CW, Negnevitsky M. Very short-term wind forecasting for tasmanian power
[38] W. Wang, C. Bruyère, M. Duda, J. Dudhia, D. Gill, M. Kavulich, K. Keene, M. Chen,
generation. IEEE Transactions on power systems 2006;21:965–72.
H.-C. Lin, J. Michalakes, S. Rizvi, X. Zhang, J. Berner, H. Soyoung, K. Fossell, ARW
[9] Monfared M, Rastegar H, Kojabadi HM. A new strategy for wind speed forecasting
Users Guide V3.9-3, 2017.
using artificial intelligent methods. Renewable energy 2009;34:845–8.
[39] Launder BE, Spalding DB. The numerical computation of turbulent flows.
[10] Jiang P, Yang H, Heng J. A hybrid forecasting system based on fuzzy time series
Computer methods in applied mechanics and engineering 1974;3:269–89.
and multi-objective optimization for wind speed forecasting. Applied energy 2019;
[40] The OpenFOAM Foundation, Openfoam v7 user guide, url:https://cfd.direct/
235:786–801.
openfoam/user-guide-v7, 2019.
[11] Liu M, Cao Z, Zhang J, Wang L, Huang C, Luo X. Short-term wind speed forecasting
[41] Liu H, Mi X, Li Y. Smart multi-step deep learning model for wind speed forecasting
based on the jaya-svm model. International Journal of Electrical Power & Energy
based on variational mode decomposition, singular spectrum analysis, lstm
Systems 2020;121:106056.
network and elm. Energy Conversion and Management 2018;159:54–64.
[12] Kani SP, Ardehali M. Very short-term wind speed prediction: A new artificial
[42] Han L, Zhang R, Wang X, Bao A, Jing H. Multi-step wind power forecast based on
neural network–markov chain model. Energy Conversion and Management 2011;
vmd-lstm. IET Renewable Power Generation 2019;13:1690–700.
52:738–45.

13
Y. Che et al. Sustainable Energy Technologies and Assessments 52 (2022) 101995

[43] Shi X, Lei X, Huang Q, Huang S, Ren K, Hu Y. Hourly day-ahead wind power [47] Bowden G, Barker P, Shestopal V, Twidell J. The weibull distribution function and
prediction using the hybrid model of variational model decomposition and long wind power statistics. Wind Engineering 1983:85–98.
short-term memory. Energies 2018;11:3227. https://doi.org/10.3390/ [48] Monaghan AJ, Barlage M, Boehnert J, Phillips CL, Wilhelmi OV. Overlapping
en11113227. interests: the impact of geographic coordinate assumptions on limited-area
[44] Dragomiretskiy K, Zosso D. Variational mode decomposition. IEEE Transactions on atmospheric model simulations. Monthly weather review 2013;141:2120–7.
Signal Processing 2014;62:531–44. https://doi.org/10.1109/TSP.2013.2288675. [49] Mahrt L. The near-surface evening transition. Quarterly Journal of the Royal
[45] Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Computation Meteorological Society 2017;143:2940–8.
1997;9:1735–80. [50] Kingma DP, Ba J. Adam: A method for stochastic optimization. 2017. <xref
[46] Delle Monache L, Nipen T, Deng X, Zhou Y, Stull R. Ozone ensemble forecasts: 2. a url=”http://arxiv.org/abs/1412.6980”><hi rend=”tt”>arXiv:1412.6980</
kalman filter predictor bias correction. Journal of Geophysical Research: hi></xref>.
Atmospheres 2006:111. [51] Taylor KE. Summarizing multiple aspects of model performance in a single
diagram. Journal of Geophysical Research: Atmospheres 2001;106:7183–92.

14

You might also like