Impact of Selected Options in the Weather Research and Forecasting Model on Surface Wind Hindcasts in Coastal Ghana

Abstract

This paper examines the impacts of five planetary boundary layer (PBL) parameterization schemes paired with several compatible surface layer (SL) parameterization schemes in the Weather Research and Forecasting Model on wind hindcasts for resource assessment purposes in a part of Coastal Ghana. Model predictions of hourly wind speeds at 3 × 3 km² and 9 × 9 km² grid boxes were compared with measurements at 40 m, 50 m, and 60 m. It was found that the Mellor-Yamada Nakanishi and Niino Level 3 (MYNN3) PBL scheme generally predicted winds with a relatively better combination of error metrics, irrespective of the SL scheme it was paired with. When paired with the Eta surface layer scheme, it often produced some of the relatively fewest errors in estimated mean wind power density (WPD) and Weibull cumulative density. A change in the simulation grid size did not have a significant impact on the conclusions of the relative performance of the PBL-SL pairs that were tested. The results indicate that the MYNN3 PBL and Eta SL pair is probably best for wind speed and energy assessments for this part of coastal Ghana.

1. Introduction

Over the years, there has been increasing interest in the use of numerical weather prediction (NWP) models, such as the Weather Research and Forecasting (WRF) model [1], for wind resource assessment. By numerically downscaling meteorological datasets, these models are used to generate wind data (wind speeds and directions) at relatively low cost for areas lacking ground measurements of such data for preliminary assessments of wind resources. Owing to diverse model options, identifying optimum model configurations (which are basically combinations of the options available in the models) for an application sometimes requires sensitivity tests, which assess, comparatively, the effects of varying model options on model performance. Predictions of surface winds by NWP models such as WRF are sensitive to model options such as simulation grid size, model physics, initial and boundary data, and parameterization of processes at the subgrid scale [2,3]. This paper focuses on selected parameterization options in the Advanced Research WRF (ARW).

Planetary Boundary and Surface Layer Parameterization

Atmospheric processes play an important role in determining certain fundamental properties of the weather and climate of the earth. Therefore, their correct representation in atmospheric models is important. In NWPs, this is done in part via the model physics, the purpose of which is to resolve and parameterize (approximate) these processes in the models [4]. Where, due to the complexity of the processes or the scales on which they occur or for other reasons, the processes cannot be explicitly represented in or resolved by the models, they are parameterized (or approximated with parameterization schemes) [4]. Parameterization involves relating the effects of such processes to variables that can more easily be determined by the models [4]. The physics parameterization schemes in WRF fall into the microphysics (MP), cumulus, long-wave radiation (Rad-L), short-wave radiation (Rad-S), land surface model (LSM), surface layer (SL), and planetary boundary layer (PBL) categories [1]. Vertical sub-grid scale transport processes in the atmosphere are parameterized by the PBL schemes, which interact directly with the SL and LSM schemes [5].

Transport processes transmit the effect of surface phenomena such as frictional drag, heat transfer, and terrain induced flow modification in the planetary boundary layer to the upper layers of the atmosphere [6]. Turbulence plays a key role in such transport processes and acts as a feedback mechanism in wind circulation [6,7,8]. PBL schemes compute turbulence flux profiles within the atmosphere, providing atmospheric tendencies of temperature, moisture, and horizontal momentum [5], which are used in predicting variables. A key difference in the PBL schemes in WRF is how they address the turbulence closure problem, which arises in the mathematical representation of turbulence (explained in several texts such as [4,9,10]), due to the difficulty of resolving the smallest turbulent eddies (which are in the order of a few milometers [6,7,8]). This is often a challenge in NWPs, as they are often run at grid resolutions that do not allow the adequate resolving of such eddies. Depending on how the closure problem is addressed in a PBL scheme, it may be classified according to an order of closure, and as a local or nonlocal closure scheme. Only vertical levels that are directly adjacent to a given point directly influence the estimation of the fluxes at that point in local closure schemes. In nonlocal closure schemes, on the other hand, multiple vertical levels influence the estimation of fluxes at a given point [11]. In addition, in WRF, most nonlocal schemes have diagnostic components for a flux profile, while the local closure schemes use turbulent kinetic energy (TKE) predicted at a point in approximating fluxes [12]. Higher order local closures and nonlocal closures are often more accurate than lower order local closure schemes [4]. Brief descriptions of several PBL schemes as well as their shortcomings have been summarized in the literature [11,12,13,14].

Parameterization methods perform differently in different atmospheric stability conditions, which inform their formulation [5,11]. Stability dependent information and other inputs needed by the PBL schemes are provided by SL schemes. The SL schemes also provide exchange coefficients for the calculation of the heat and moisture fluxes by LSM schemes. These fluxes serve as bottom boundary conditions for the PBL schemes [15,16]. Key differences among SL schemes include the approaches and methods used in computing surface exchange coefficients [5]. However, they are mostly based on the similarity theory, which is explained in texts such as [6,9,10]. In WRF, PBL schemes are recommended to be used with specific SL schemes but are generally compatible with most of the LSM schemes in the model.

Given the importance of wind speeds to wind energy extraction, and as wind turbines operate in the lower parts of the PBL, several studies [3,17,18,19,20,21,22,23] over the years have examined the impact of PBL schemes in WRF on the wind hindcasts. However, studies in the tropics [17,18,19,20,21] have often not tested the different PBL schemes with different compatible SL schemes. In addition, the impact of the schemes on model performance is influenced by local terrain features and atmospheric conditions, which often vary with geographical location [2,3]. Against this background, in this paper, we investigate the impact of selected PBL schemes, paired with several compatible SL schemes, on wind hindcasts for wind energy assessment purposes in an area in coastal Ghana. The study focuses on five PBL schemes selected from a preliminary study of PBL schemes in coastal Ghana, and other studies in tropical areas [17,18,19,20,21]. These are:

• 1st order hybrid (local/nonlocal) closure Asymmetric Convective Model (ACM2) [24]
• 2nd order TKE closure Mellor-Yamada Nakanishi Niino Level 3 (MYNN3) [25]
• 1.5 order TKE closure University of Washington (UW) [26]
• 1.5 order TKE closure Grenier-Bretherton-McCaa (GBM) [27]
• 1st order nonlocal closure Yonsei University (YSU) [28]

The aim of the study is to offer some insight into the relative impacts of the selected PBL-SL pairs on wind speed and mean wind power density estimates by the model for coastal Ghana. The rest of the paper is organized as follows; Section 2 covers the study area, verification data, model configuration, and experimental design. Section 3 presents and discusses results of analysis, and Section 4 summarizes the study and presents conclusions drawn from the study.

2. Materials and Methods

2.1. Study Area and Data

The study area covers the coastal plains of South East Ghana (shown in Figure 1). The area comprises predominantly low-lying coastal plains with savanna grass vegetation and experiences two main seasons in a year: a harmattan season that is dominated by dry and dusty desert winds from the North-East, starting from around November and lasting until February, and a bimodal rainy season dominated by Monsoon winds that ends around November [29,30]. The Energy Commission of Ghana (EC) has conducted mast measurements at selected sites, mostly along the coast of this region. The observed (measured) data for this study, which comprise hourly measurements of wind speeds (in selected months) in 2013, at heights of 40, 50, and 60 m above ground level, is from one such EC masts, located at 5.7861 °N and 0.9188 °E.

2.2. Model Configuration

Version 3.8.1 of the Advanced Research WRF (ARW) [1] was used for this study. Key features of the model include a fully compressible, non-hydrostatic Euler equation, a terrain following a vertical coordinate system, and a staggered horizontal grid. Model prognostic variables include three-dimensional wind, turbulent kinetic energy, and potential temperature. Detailed descriptions of the model physics, equations, and dynamics are provided by [1].

The model configuration for the study is summarized in

Table 1. Model configuration.

Model Version	Advanced Research WRF v3.8.1
Initial and Boundary Conditions	NCEP Final Analysis (GFS-FNL) [35]: 1° × 1° and 6 h Resolution
Land Use Data	30-arc-second USGS1 with lakes
Topographical Data	30-arc-second USGS GMTED2010
Map Projection	Mercator
Vertical Resolution	40 vertical pressure levels (automatically set)
Horizontal Resolution (km)	27	9	3
Domain Size (grid points)	91 × 103	82 × 94	64 × 55
Model Timestep (seconds)	120
FDDA2	Analysis Nudging (Disabled in the PBL)
Parameterization Schemes:
Cloud Microphysics (MP)	Eta microphysics [36]
Long-Wave Radiation (LW-Rad)	Rapid Radiative Transfer Model [37]
Short-Wave Radiation (SW-Rad)	Dudhia [38]
Surface Layer (SL)	i. Mellor-Yamada Nakanishi Niino (MYNN) ii. Pleim-Xiu (PX) [39] iii. Revised MM5 Similarity (R-MM5) [40] iv. Eta Similarity (Eta) [41,42,43]
Land Surface Model (LSM)	Unified Noah [44] Pleim-Xiu (PX) [45,46]
Planetary Boundary Layer (PBL)	i. ACM2 [24] ii. GBM [27] iii. MYNN3 [25] iv. UW [26] v. YSU [28]
Cumulus	Kain-Fritsch [47] (turned off for domain 3 [1,32])

¹ United States Geological Survey. ² Four-Dimensional Data Assimilation.

. Map projections transform atmospheric properties (defined on earth’s spherical surface) to a flat model grid [4] to enable the application of grid point methods to solutions of the atmospheric flow equations. Map projections tend to affect model stability as they distort distances at any given point, affecting the maximum stable timestep in the WRF solver. To maintain numerical solution stability, it is recommended to use a projection that keeps the map-scale factor (a measure of distance distortions from the transformation) close to unity over the simulation grid [4]. For low-latitude and tropical regions, the Mercator projection is recommended as it best satisfies this (stability) condition [4,31]. To further ensure model stability, a model timestep of 120 s (less than the maximum 6 times the magnitude of the coarsest horizontal grid distance) was used as suggested by [1]. The domains, shown in Figure 2, have horizontal resolutions of 27 km, 9 km, and 3 km, and a vertical resolution of 40 vertical pressure levels each. The horizontal resolutions were chosen to achieve a nesting ratio of 3, and the final horizontal resolution of 3 km was used because it was found to be optimal for wind simulations in WRF [32,33]. The vertical resolution was chosen following recommendations of [34]. The model top was 50 hPa with the lowest half level at approximately 28 m asl.

The selected PBL parameterization schemes were paired with compatible SL schemes as recommended by [31,48] (except the old MM5 scheme). Other required parameterization schemes were selected based on other wind sensitivity studies in coastal Ghana [49,50], and practices from other wind sensitivity studies (mostly in the tropics) [17,18,19,20,21]. The Eta Microphysics, New Rapid Radiative Transfer Model [37] and Dudhia [38] schemes were used for MP, LW-Rad, and SW-Rad parameterizations, respectively. Cumulus parameterization was not used in domain 3, as the horizontal grid resolution in this domain was considered fine enough for adequate resolving of cumulus processes [1,32]. For domains 1 and 2, however, cumulus processes were parameterized with the updated Kain-Fritsch scheme [47]. The Unified Noah LSM [44], was used for land surface parameterization. In addition, following recommended best practices on the use of the ACM2 PBL scheme and PX SL schemes [51], the PX LSM [45,46] was also tested but with the PX SL scheme only. The resulting PBL-SL-LSM configurations that were tested are presented, with references as obtained from the WRF physics page, in

Table 2. Configurations tested.

No.	Designation	PBL Scheme	SL Scheme	LSM Scheme
1	ACM2-P-P	ACM2	PX	PX
2	ACM2-P-N	ACM2	PX	Noah
3	ACM2-R-N	ACM2	R-MM5	Noah
4	GBM-R-N	GBM	R-MM5	Noah
5	MYNN3-M-N	MYNN3	MYNN	Noah
6	MYNN3-R-N	MYNN3	R-MM5	Noah
7	MYNN3-E-N	MYNN3	Eta	Noah
8	UW-R-N	UW	R-MM5	Noah
9	UW-E-N	UW	Eta	Noah
10	YSU-R-N	YSU	R-MM5	Noah

.

2.3. Experimental Design

A total of 10 configurations were tested. Owing to limited computational resources, each configuration was used to simulate a period comprising four months, January, February, May, and September of 2013, one month at a time. The months were selected for their relatively high or low monthly average wind speeds in the seasons that pertained in this part of Ghana; January and March represented the harmattan season and May and September, the rainy season. It was hoped that by selecting the periods simulated in this manner (as has been done in other resource assessment sensitivity studies [22,32]), the effect of major seasonal changes on annual winds would be captured by the options being tested. The grid nudging option of the WRF Four-Dimensional Data Assimilation (FDDA) system is a technique that has been used in several studies on wind downscaling for resource assessment purposes [20,52,53]. The technique bridges the gap between the model simulations and time-interpolated values from input data. All three simulation domains were nudged during all the simulations, following practices of previous studies [20,52,53]. Nudging options and simulation lengths were chosen based on recommendations from a previous study in coastal Ghana [49].

2.4. Postprocessing of Data and Evaluation of Options

Post-processing of results generally followed the procedure used in previous studies in the study area [49,50]. However, hourly predictions of winds (at 10 m and other relevant half vertical levels) were bilinearly interpolated to the mast location. Furthermore, winds were interpolated to the heights of analysis (40 m, 50 m, and 60 m), with log-linear interpolation [54]. Scheme performance was assessed in terms of four error metrics, which were calculated with procedures from previous studies [49,50]: mean error (ME), root mean square error (RMSE), standard deviation of the error (STDE), as well as the correlation coefficient (CC) of the predictions. The error metrics were combined into a prediction skill score (SS) (as was done in previous studies [49,50]), which was used to rank the options. In addition, the ME, RMSE, and CC were compared to values that were considered as indicators of good model performance in studies [18,19,55].

As the intended application of the findings of the study is wind energy assessment, the impacts of the options on wind power estimation were evaluated by comparing their Weibull cumulative distributions and mean wind power densities to those from observations for the study period. The Weibull distribution is widely used in many fields of the wind energy industry. The cumulative distribution gives the probability of wind speeds being less than or equal to the speed at which it is evaluated. Its function is given as [56]

$$F(v)=1-\exp \left[-{\left(\frac{v}{c}\right)}^k\right]$$

(1)

where v, c, and k are the wind speed, Weibull scale, and shape factors, respectively. The scale and shape parameters were estimated using the empirical (mean and standard deviation) method (with formulas from [56]). This method was chosen after an evaluation of five methods—the empirical, moment, graphical or least squares, the energy pattern factor, and maximum likelihood methods—using an evaluation method from a study by [56]. The empirical method was chosen for its simplicity and often relatively better or on-par relative test rank, which was in terms of total normalized results.

The distributions for observed data and the configurations were compared via the maximum absolute error of the cumulative distribution function (max CDF error), which was determined as the maximum difference between the cumulative distributions of observed and predicted data, (evaluated in 0.5 m/s bins as recommended by [57]) [53]:

$$\mathrm{Max}\mathrm{CDF}\mathrm{Error}=\max \left|F{(v_i)}_{obs}-F{(v_i)}_{sim}\right|.$$

(2)

The mean wind power densities were expressed as the percent error of the difference between the mean WPD of observed and predicted data for a period of evaluation. The error was then expressed as a percentage of the mean WPD from observations. The mean WPDs were determined from the estimated Weibull parameters as [56,57]

$$\mathrm{WPD}=\left[\frac{1}{2}\rho c^3\mathsf{\Gamma }\left(1+\frac{3}{k}\right)\right]$$

(3)

where $\rho$, the air density, was assumed to be 1.160 kg m⁻³, as estimated in a previous study [49] in the study area.

3. Results

Results at 60 m for the PBL-SL pair options are presented in

Table 3. Error metrics and skill scores at 60 m for study period.

	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	Weibull k	Weibull c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|
Observation	5.87						2.93	6.58	167
ACM2-P-P	5.83	−0.04	1.65	1.65	0.67	0.3	3.78	6.59	145	−13.1	0.0645
ACM2-P-N	5.95	0.08	1.66	1.66	0.67	0.2	3.78	6.59	154	−7.6	0.0693
ACM2-R-N	6.08	0.21	1.65	1.64	0.67	1.0	3.86	6.72	163	−2.2	0.0888
GBM-R-N	6.06	0.19	1.60	1.59	0.69	2.0	4.04	6.68	158	−5.1	0.0964
MYNN3-M-N	5.68	−0.19	1.55	1.54	0.71	3.3	3.56	6.31	137	−17.6	0.0731
MYNN3-R-N	5.71	−0.16	1.55	1.54	0.72	3.5	3.34	6.36	143	−14.2	0.0545
MYNN3-E-N	5.96	0.09	1.58	1.57	0.72	2.6	3.33	6.64	163	−2.2	0.0399
UW-R-N	5.86	−0.01	1.57	1.57	0.70	2.1	3.67	6.49	149	−10.9	0.0530
UW-E-N	6.19	0.32	1.63	1.60	0.70	2.2	3.70	6.85	174	4.6	0.0932
YSU-R-N	6.01	0.14	1.60	1.59	0.69	1.9	3.85	6.64	157	−5.7	0.0802

. It can be seen from the error metrics that the impacts of all the tested PBL-SL pairs on model performance were mostly within the acceptable limits: RMSE < 2 m/s, ME < 0.5 m/s. However, the ACM2, GBM, and YSU predictions had CCs that were slightly less than the acceptable limits (CC ≥ 0.7 [18,19,55]). It can also be observed that for the same PBL scheme, although the choice of an SL scheme did have some impact on the average wind speed prediction, the impact (if any at all) was less on the error metrics. Again, for the same PBL scheme, the Eta SL scheme produced higher average wind speeds than the other SL schemes. The YSU PBL scheme predicted higher average wind speeds than the MYNN3 PBL scheme (irrespective of SL scheme). Generally, configurations with the MYNN3 PBL scheme predicted with some of the least RMSEs, the best consistency (least STDEs), and highest correlations, and therefore, gave the best skill scores, although as has been mentioned, the error metrics for all the other configurations (except for the CCs of the ACM2, GBM, and YSU PBL schemes) were within acceptable limits for good model performance. In contrast to the average wind speed predictions, the configurations had more significant impacts on WPD estimates. However, some of the trends observed in the speed predictions were also observed in the WPD estimates; the configurations with the Eta SL scheme gave relatively smaller absolute WPD errors than the R-MM5 SL scheme. The MYNN3-E-N and ACM2-R-N configurations gave the best average WPD estimates for the entire study period. However, as can be seen from the table, the MYNN3-E-N configuration had a better maximum error of CDF. In addition, it can be observed from the CDF plots presented in Figure 3 that the probability plot of the MYNN3-E-N configuration was closest to the plot of observed data for speeds below 7 m/s; the other options gave relatively lower probabilities. For higher speeds, the MYNN3-E-N together with the UW-E-N configurations gave the closest probability plots.

Similar trends were often observed when the analysis was restricted to the seasons in the area; a change in SL scheme did not often produce a significant change in metrics. The Eta SL scheme produced higher average wind speed estimates than the R-MM5. The YSU scheme simulated higher wind speeds than the MYNN3 scheme. Configurations with the MYNN3 PBL scheme still ranked best in the rainy season and all the options satisfied all the criteria for good performance (except the ACM2 configurations, which had CCs < 0.7 in the rainy season). Notable exceptions to these trends are the ACM2 PBL scheme with the Noah LSM (with either SL schemes) ranking relatively better for speed prediction during the harmattan season. In addition to the MYNN3-E-N configuration, the UW-R-N configuration gave a relatively good WPD error with a better max CDF error in the harmattan season. Results on the seasonal analyses at 60 m are available in

Table A2. Error metrics skill scores and WPD for PBL-SL pairs at 60 m (for seasons).

	Rainy Season											Harmattan Season
	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	k	c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	k	c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|
Observation	6.03						2.77	6.77	187			5.70						3.18	6.37	147
ACM2-P-P	5.77	−0.26	1.82	1.80	0.64	1.6	4.27	6.42	135	−27.9	0.0773	5.90	0.19	1.44	1.43	0.73	1.6	3.43	6.76	157	7.0	0.0468
ACM2-P-N	5.84	−0.19	1.85	1.84	0.62	0.7	4.37	6.42	139	−25.7	0.0880	6.07	0.37	1.43	1.38	0.75	3.5	3.40	6.76	171	17.0	0.0769
ACM2-R-N	6.04	0.01	1.83	1.83	0.63	0.2	4.43	6.63	153	−18.3	0.1105	6.13	0.43	1.44	1.38	0.75	3.6	3.46	6.82	175	19.6	0.0897
GBM-R-N	5.98	−0.05	1.75	1.75	0.67	1.7	4.31	6.58	150	−19.7	0.0989	6.13	0.43	1.44	1.38	0.73	3.0	3.82	6.79	168	14.9	0.1056
MYNN3-M-N	5.85	−0.18	1.69	1.68	0.70	3.1	4.00	6.47	145	−22.7	0.0713	5.50	−0.20	1.40	1.38	0.74	2.9	3.22	6.15	131	−10.6	0.0000
MYNN3-R-N	5.91	−0.11	1.65	1.65	0.72	3.4	3.78	6.56	152	−18.8	0.0638	5.49	−0.21	1.43	1.42	0.74	2.4	3.01	6.15	135	−8.0	0.0000
MYNN3-E-N	6.18	0.15	1.68	1.67	0.71	3.2	3.78	6.85	173	−7.5	0.0900	5.73	0.02	1.47	1.47	0.74	1.2	2.98	6.42	154	5.0	0.0183
UW-R-N	5.86	−0.17	1.71	1.70	0.69	2.9	3.86	6.46	145	−22.5	0.0623	5.88	0.17	1.42	1.41	0.73	2.1	3.49	6.54	154	5.0	0.0468
UW-E-N	6.19	0.16	1.72	1.71	0.69	2.6	3.90	6.85	172	−8.1	0.0982	6.18	0.48	1.54	1.46	0.72	1.0	3.52	6.88	179	21.9	0.1016
YSU-R-N	5.99	−0.04	1.74	1.74	0.68	1.8	4.36	6.58	150	−19.9	0.1015	6.03	0.33	1.44	1.40	0.74	2.6	3.48	6.71	166	13.5	0.0724

in the Appendix A.

Changes in heights of analysis and the simulation grid box sizes did not have significant impacts on most of the above trends either. At the lower heights, the relative performances of the configurations for speed predictions at the lower heights were also largely the same. However, the UW-E-N and GBM-R-N configurations tended to give relatively better mean WPD errors, although the max CDF error of the MYNN3-E-N configuration was still relatively better. In the seasonal analyses, the MYNN3-E-N still tended to give some of the relatively best (if not the best) mean WPD and max CDF error. Data from the 9 × 9 km² grid produced generally lower average wind speeds and general increase in absolute ME with little impact on the scheme rankings presented earlier. Max CDF errors were also higher when compared to those estimated with data from the 3 × 3 km² grid. The MYNN3-E-N and UW-E-N configurations still tended to give relatively lower mean WPD errors with the max CDF error of the former better as compared to the other options. Selected results on these analyses are available in

Table A1. Error metrics and skill scores for PBL-SL pairs at 50 m and 40 m (for entire study period).

	50 m											40 m
	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	k	c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	k	c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|
Observation	5.78						2.91	6.48	160			5.68						2.89	6.37	153
ACM2-P-P	5.53	−0.25	1.66	1.64	0.66	0.8	3.81	6.26	124	−22.6	0.0413	5.32	−0.36	1.68	1.64	0.65	0.9	3.84	6.04	110	−27.9	0.0360
ACM2-P-N	5.66	−0.12	1.65	1.65	0.65	0.3	3.82	6.26	132	−17.1	0.0527	5.46	−0.22	1.67	1.65	0.64	0.5	3.86	6.04	118	−22.4	0.0480
ACM2-R-N	5.81	0.03	1.63	1.63	0.66	0.7	3.97	6.42	142	−11.3	0.0762	5.63	−0.05	1.62	1.62	0.66	0.8	4.04	6.21	128	−16.4	0.0743
GBM-R-N	5.86	0.08	1.58	1.58	0.68	1.9	4.03	6.47	144	−9.7	0.0858	5.73	0.05	1.58	1.58	0.68	1.9	4.04	6.33	135	−11.6	0.0845
MYNN3-M-N	5.40	−0.37	1.58	1.53	0.71	3.4	3.71	5.99	117	−27.0	0.0270	5.21	−0.47	1.60	1.53	0.70	3.2	3.83	5.77	103	−32.2	0.0295
MYNN3-R-N	5.50	−0.28	1.54	1.52	0.72	3.7	3.47	6.13	127	−20.6	0.0199	5.36	−0.32	1.54	1.51	0.71	3.7	3.55	5.96	116	−24.0	0.0242
MYNN3-E-N	5.71	−0.06	1.55	1.55	0.71	2.7	3.40	6.37	143	−10.3	0.0318	5.55	−0.13	1.54	1.54	0.71	2.9	3.46	6.18	130	−14.7	0.0324
UW-R-N	5.66	−0.11	1.57	1.57	0.69	2.3	3.77	6.28	133	−16.4	0.0504	5.53	−0.15	1.58	1.57	0.68	2.2	3.84	6.12	123	−19.1	0.0525
UW-E-N	5.95	0.17	1.59	1.59	0.69	2.1	3.75	6.60	155	−2.7	0.0804	5.79	0.11	1.59	1.58	0.68	2.0	3.77	6.42	143	−6.3	0.0765
YSU-R-N	5.79	0.02	1.57	1.57	0.69	1.9	3.96	6.41	140	−12.0	0.0741	5.65	−0.03	1.56	1.56	0.69	2.1	4.04	6.24	129	−15.3	0.0765

and

Table A3. Error metrics skill scores and WPD for PBL-SL pairs for entire study period at 60 m (9 × 9 km² grid).

	Average Wind Speeds (m/s)	ME (m/s)	RMSE (m/s)	STDE (m/s)	CC	Skill Score	k	c	Mean WPD (Wm−2)	WPD Error (%)	Max |CDF Error|
Observation	5.87						2.93	6.58	167
ACM2-P-P	5.62	−0.25	1.68	1.66	0.66	0.6	3.63	6.38	132	−20.8	0.4392
ACM2-P-N	5.75	−0.12	1.66	1.66	0.66	0.6	3.69	6.38	141	−15.4	0.1537
ACM2-R-N	5.89	0.02	1.64	1.64	0.67	0.9	3.75	6.53	151	−9.7	0.1374
GBM-R-N	5.87	0.00	1.59	1.59	0.69	1.9	3.96	6.49	146	−12.5	0.1027
MYNN3-M-N	5.45	−0.42	1.61	1.56	0.71	3.3	3.47	6.07	123	−26.2	0.7957
MYNN3-R-N	5.50	−0.37	1.58	1.54	0.72	3.8	3.27	6.14	130	−22.2	0.7249
MYNN3-E-N	5.77	−0.10	1.58	1.57	0.72	2.9	3.22	6.45	151	−9.3	0.1606
UW-R-N	5.69	−0.17	1.59	1.58	0.70	2.5	3.61	6.33	138	−17.3	0.2823
UW-E-N	6.05	0.18	1.61	1.60	0.69	2.1	3.66	6.72	165	−1.2	0.4565
YSU-R-N	5.81	−0.06	1.61	1.61	0.68	1.7	3.82	6.43	143	−14.2	0.0372

in the Appendix A.

4. Discussion

The performance of PBL and SL schemes differ in different atmospheric stability conditions, which inform the methods used by the schemes [58]. Due to inadequate data, we were not able to assess stability conditions in the study area, and so are unable to assess the performance of the schemes against some of these conditions. However, we find that our observations are consistent with results from several studies. For instance, the YSU scheme was observed to generally predict higher wind speeds than the MYNN3 scheme due the relatively shallower mixed layer that it simulates [2,3,21]. PBL schemes were found to have more significant impacts on error metrics of surface wind speed hindcasts than other parameterization schemes in studies in similarly coastal terrains [2,22]. Furthermore, the often relatively better ranking of the MYNN3 PBL scheme (over the other local closure schemes) is possible due to it being of a higher order closure. Higher order local (and nonlocal) closure schemes generally yield more accurate results than schemes that employ lower order local closures [4].

The diurnal profiles of average wind speeds for the seasons in the area (shown in Figure 4) are also consistent with what has been reported from other sensitivity studies of PBL schemes in WRF to wind hindcasts in tropical and coastal areas [3,19,21]. Similar peak winds and a relatively high overestimation of winds around and after sunset have been reported for a tropical area. In addition, winds peaking between noon and sunset were also reported in a relatively cooler season in the same area [19]. It was explained that the overestimation of winds after sunset can be attributed to the inability of the PBL schemes to decouple air near the surface and aloft at night, as a result of differences in vertical mixing strength and entrainment of air above the PBL [3,21].

The profiles also suggest local winds to be from land-sea circulations. Land and sea breezes result from a convective cycle where warm air over land rises to be replaced by cool sea breezes during the day. The cycle reverses at night as land cools more rapidly than the sea. The Land-sea temperature difference plays a key role in the strength of this cycle, often producing stronger sea breezes (Figure 4), as it is higher during the day [59]. The profiles and reports on the relative predictions of temperature by the Eta and R-MM5 SL scheme also offer possible explanations as to why winds produced by configurations with the Eta scheme are higher than those with the R-MM5. Higher land temperatures during the day should result in higher land-sea temperature differences (as the sea temperature rises at a relatively lower rate) and thus stronger winds [59]. The Eta SL scheme was reported to produce relatively higher temperatures (due to its higher heat fluxes and exchange coefficients) than the R-MM5 [60] (citing [61,62]). The diurnal plots of the average temperature at 2 m (T2) shown in Figure 5, which are consistent with those reported by [19] and [60] (citing [62,63,64,65]), suggests this to be the case in the area. In addition, we see from the profiles for both seasons that the MYNN3-M-N configuration, which often predicted some of the lowest average wind speeds (in both seasons), also recorded some of the lowest T2s diurnally. However, despite the relatively higher temperatures in the harmattan season, relatively lower average wind speeds were observed during this season. This is possibly due to a weakening of the breezes by the northeast winds that blow during the harmattan.

5. Summary and Conclusions

Though the Mesoscale Atmospheric Simulation System (MASS) NWP model from the AWS Truewind MesoMap system has been used to assess Ghana’s wind resources in the past, it is a propriety model. In addition, there was a lack of adequate mast measurements at the time limited verifications and adjustments of the model for optimum performance over coastal Ghana [66]. The open source nature and increasing popularity of WRF for similar purposes makes it an attractive alternative for future assessments. Predictions of surface winds by WRF are sensitive to model options such as physics, simulation grid size, and parameterization of processes at the subgrid scale. In this paper, we tested the sensitivity of wind in coastal Ghana to five planetary boundary layer (PBL) parameterization schemes (selected based on a preliminary study and other studies [17,18,19,20,21,50,67]), paired with different compatible surface layer (SL) schemes [31,48]).

It was found that hindcasts from all 10 PBL-SL pairs generally had speed prediction error metrics within or close to acceptable limits for good performance, as established from other sensitivity studies (RMSE < 2 m/s, ME < 0.5, CC > 0.7 [18,19,55]). However, they differed in their prediction of the mean wind power densities (WPD) and cumulative distribution functions for the period, in consistency and accuracy. Hindcasts with the MYNN3 PBL scheme generally had a relatively better combination of error metrics, and when combined with the Eta SL scheme, it often gave the best or some of the best WPD and maximum errors of CDF. Relative to the other SL schemes, the Eta SL scheme tended to predict relatively higher wind speeds for the same PBL scheme. At lower heights, the UW and GBM PBL schemes (with the Eta and R-MM5 SL schemes, respectively) tended to give better mean WPD errors, but the MYNN3 with the Eta SL scheme still gave better skill scores and max CDF errors. The above trends were also largely observed in the two seasons that pertain along the coast of Ghana with few exceptions. A change in grid resolution was not found to significantly affect the trends in the relative performance of the options.

Though we were not able to assess the performance of the schemes against different atmospheric conditions, several of the trends from our results were found to be consistent with what has been reported by other studies in the literature, based on which we believe our other observations and conclusions are largely credible. Some of such results from other studies include the following: a change in SL scheme did not have significant impacts on most of the error metrics [2,22]; the YSU PBL scheme simulated higher winds than the MYNN3 PBL scheme [2,3,21]; average wind speeds between sunset and sunrise were overpredicted [3,21]; the Eta SL scheme predicted higher T2s than the R-MM5 scheme [60,63]; and no one option was always superior to the others [2,11].

The MYNN3-E-N configuration (as tested in this study with the GFS FNL) is most consistent in predicting with relatively better combined wind speed error metrics and errors of CDF, as well as relatively good mean WPD errors with changing factors (i.e., height, simulation grid box size, and seasons in coastal Ghana). In addition, when predictions from this configuration (MYNN3-E-N) for the study period were compared with monthly average wind speeds for four other locations along the coast of Ghana estimated from [68], results were largely reasonable; average mean error was within 0.5 m/s for the locations (see

Table A4. Comparison of monthly average wind speeds at 60 m for other locations in coastal Ghana (observations were estimated from [68]).

	January	February	May	September	Average
SEGE (5.872 °N, 0.345 °E)
WRF	4.91	5.76	4.86	6.49	5.55
Observations	4.6	5.7	4.1	6.6	5.25
Error	0.31	0.06	0.76	−0.11	0.25
DZITA (5.774 °N, 0.714 °E)
WRF	4.93	6.05	4.84	6.88	5.68
Observations	5	6.1	4.5	7.3	5.73
Error	−0.07	−0.05	0.34	−0.42	−0.05
MANKOADZE (5.317 °N, 0.70 °W)
WRF	4.56	5.11	4.24	6.20	5.03
Observations	4.5	5.3	4.5	5.4	4.93
Error	0.06	−0.19	−0.26	0.8	0.1
DENU (6.112 °N, 1.141 °E)
WRF	5.00	6.01	4.89	6.27	5.54
Observations	4.5	5.5	4	6.2	5.05
Error	0.50	0.51	0.89	0.07	0.49

* WRF speeds are from the MYNN3-E-N hindcasts (on the 9 × 9 km² grid box).

in the Appendix A). Based on this we conclude that the MYNN3-E-N configuration could be considered suitable for wind hindcasts for resource assessments in coastal Ghana.