Wind Energy Assessment in Forest Areas Using Multi-Source Optimized WRF Model

Abstract

Accurate wind field simulation in forest areas is crucial for wind energy development but remains challenging for traditional WRF models due to complex terrain and vegetation heterogeneity. This study proposes a multi-source optimization framework integrating seasonal PBL scheme selection, localized leaf area index (LAI) adjustment, and 3DVAR data assimilation to improve WRF performance in forested terrain. The framework was validated using observations at 20 m, 50 m, and 100 m heights in Maoershan forest area. Results show that: (1) PBL schemes exhibit significant seasonal dependence—YSU performs best in spring (unstable conditions), while MYJ shows slight advantages near the surface in winter (stable conditions). (2) Localized LAI correction reduces near-surface wind speed bias by 35% and improves wind direction accuracy by 28%, with stronger effects in summer. (3) 3DVAR assimilation further enhances accuracy, achieving correlation coefficients of 0.869 for wind speed and 0.813 for wind direction, with greater improvements in summer and near the surface. (4) Winter wind power density at 100 m reaches 475 W/m², 38% higher than summer, indicating stable exploitable resources. The proposed framework provides a replicable methodology for wind field simulation in forest regions worldwide.

1. Introduction

Countries around the world are facing the dual challenges of the depletion of natural resources and the continuous deterioration of the ecological environment. As a renewable energy source, wind energy features low overall development and utilization costs and minimal environmental pollution. As the world’s largest wind energy market, China’s new wind power installed capacity accounted for 70% of the world’s new wind power installed capacity in 2024, and wind energy applications are increasingly popular. [1,2]. As the country’s exploitable wind resources in plains and offshore regions approach saturation [3], attention has increasingly shifted toward forested regions [4]. According to national statistics on the overlap between forested areas and wind resource coverage, forested areas account for 23.46% of China’s total land area [5]. Among these, the northeastern forest regions represent nearly one-third of China’s total forested area [6], indicating considerable untapped potential for wind energy development [7]. A comprehensive understanding and accurate assessment of the wind potential in candidate regions are essential prerequisites for rational wind farm planning and efficient utilization of wind energy resources [8].

The development of wind energy in forest areas is not unique to China, but in the general trend of the global transition to renewable energy, forest areas have gradually become an important location direction for onshore wind power development [9]. Nordic countries with high forest cover are increasingly locating wind farms in such areas [10]. However, the development of wind energy resources in forest areas faces a double challenge. On the one hand, the wind field in forest areas has a significantly high complexity [11]. Due to the joint influence of rolling terrain and dense vegetation, the near-earth atmospheric boundary layer air flow undergoes deflection, separation and abrupt change, wind direction disturbance is frequent, and the vertical and horizontal exchange of momentum, heat and water vapor between atmospheric turbulence and the boundary layer is more complex [12,13,14]. The complex tree crown structure (characterized by tree height and leaf area index) will cause high intensity turbulence and strong vertical wind shear [15], which brings great uncertainty to the analysis of wind field characteristics and wind energy assessment in forest areas [16]. On the other hand, seasonal changes in vegetation (such as defoliation and foliation) will cause dynamic changes in surface roughness, further increasing the complexity of wind resource assessment [17]. Some researchers have found [18,19] that the prediction error of near-surface wind speed of wind farms in China is about 25% to 40%, especially in mountainous areas with complex terrain and significantly affected by ground fluctuations and vegetation cover changes. Wang et al. [20] used the WRF model to simulate the near-surface wind field of complex terrain. The results showed that although different PBL schemes could accurately predict the dominant wind direction and wind speed trend, there were still systematic deviations. Therefore, it is very important to fully understand and accurately assess the wind energy potential of forest areas [21], and it is necessary to explore an accurate method for better prediction and analysis of wind resources [22].

Assessing wind energy potential over complex terrain requires detailed statistical analysis of wind resource data. Currently, three primary methods are used to obtain wind resource data [23]: field measurements [24,25], wind tunnel experiments [26,27], and numerical simulations [28,29]. Field measurements can provide spatiotemporal wind data under real atmospheric conditions, which is suitable for studying large-scale wind field characteristics and their ecological effects [30]. However, it is limited by the high cost and the limited number of wind measuring towers [31,32,33]. Wind tunnel experiment, on the other hand, can provide controllable experimental conditions and high-resolution local wind field data [34], yet it does not have obvious advantages due to the consideration of cost, cycle and test accuracy [35,36]. The numerical simulation method can efficiently construct regional wind field models under real terrain conditions and has significant advantages when combined with reanalysis meteorological data [35]. By configuring various parameter schemes, these methods can adapt to different simulation processes [37,38]. With the rapid advancement of computer technology, their applications have become increasingly widespread and mature. Therefore, this study employs numerical simulation to acquire wind field data and conduct corresponding wind energy assessments.

In the field of wind energy, the Weather Research and Forecasting (WRF) model is the most widely used numerical simulation tool [35,39,40]. The WRF model overcomes the dependence on in situ meteorological measurements such as mast data and features a well-developed dynamic framework with diverse physical parameterization schemes. Its open-source structure allows for effective simulation and assessment of wind fields [41]. However, as a mesoscale model with a spatial resolution typically on the order of kilometers, WRF’s predictive accuracy remains limited. What is more, the vertical and horizontal exchange of momentum, heat, water vapor and so on in atmospheric turbulence and the atmospheric boundary layer is more complex [12,13,14], which makes it difficult to fully meet the actual needs of wind power development. To improve simulation precision, researchers have commonly enhanced terrain resolution [42], refined land-use categories [43], or tested various combinations of physical parameter schemes [21]. Updating the default WRF sea surface temperature (SST) and roughness length (Z₀) significantly improved wind simulation accuracy in coastal regions, with greater enhancement near the coastline than offshore [44]. In forested ecological regions, wind fields are further influenced by heterogeneous vegetation cover and complex canopy structures [45]. The leaf area index (LAI), a key parameter describing canopy morphology, directly affects turbulent transport rates in the near-surface layer and consequently modulates momentum, mass, and energy exchanges between the surface and the atmosphere [46]. It is therefore considered a primary factor governing vertical wind profiles over vegetated areas [47]. The default Noah land surface scheme of WRF uses typical climatological values to set LAI, which cannot fully reflect the local canopy characteristics. Taking the study of wheat dry-hot wind as an example, Wang Shu et al. found that PKULM was more accurate in expressing vegetation physiological processes due to the introduction of higher time resolution LAI data. Finally, the latent heat flux simulated by PKULM is more consistent with the physiological process law and observation law of vegetation than Noah LSM [48]. Hence, examining the impact of modified LAI values on wind field simulations is of significant importance for improving model performance in forested terrains.

The accuracy of the atmospheric initial state is another critical factor influencing WRF simulation performance [49]. Fu et al. [50] compared simulations of high-wind typhoon events driven by two reanalysis datasets, NCEP-FNL and ECMWF-ERA5, and found that higher-resolution forcing data improved both the representation of typhoon structure and the accuracy of peripheral wind fields. Yang et al. [51] found that the WRF model could reproduce the mean climate, interannual variations, precipitation cycles, and surface temperature (T2 m), though with some deviations. Shi et al. [52] reported that assimilating observational data into the initial atmospheric fields markedly improved model quality and simulation accuracy. Commonly used data assimilation techniques include three-dimensional variational method (3DVAR), four-dimensional variational method (4DVAR), ensemble Kalman filter method (EnKF), and variational assimilation method (Hybrid) [53]. Wang et al. [54] compared precipitation simulations with and without four-dimensional nudging and observed significant improvements in precipitation distribution and magnitude after assimilation. Among these approaches, the 3DVAR method remains the most widely applied. In addition, Yang et al. [55] investigated the benefits of improving surface information quality to enhance rainfall simulation, using the enhanced land use and leaf area index (LAI) dataset from the WRF-3 DVAR assimilation system as a case study.

It is worth noting that the aforementioned improvements mainly focused on a single factor and lacked collaborative optimization studies on multi-source information such as parametric schemes, surface parameters, and initial fields. Especially in forest areas, studies that comprehensively consider the seasonal applicability of PBL schemes, local adjustments of canopy parameters, and data assimilation are still rarely reported. To address these research gaps, this study proposes the following core hypotheses: By integrating the multi-source collaborative optimization framework of PBL scheme seasonal selection, LAI local adjustment, and 3DVAR data assimilation, the wind field simulation accuracy in forest areas can be significantly improved, and there are certain physical laws. To verify this hypothesis, this study designs four research questions: (i) Are there significant differences in the performance of different PBL schemes in different seasons and at different heights in the northeastern forest area? What are the patterns? (ii) Can the LAI correction based on local data effectively reduce the systematic deviation in the simulation of near-ground wind speed? What are the differences in the improvement effect at different heights? (iii) Can the introduction of 3DVAR data assimilation further improve the wind field simulation accuracy? (iv) How does the wind field simulation result after multi-source collaborative optimization affect the assessment of wind energy resources? What is the optimization effect? To answer these questions, the main research activities include: (i) constructing simulation experiments of PBL schemes at different heights in spring and winter; (ii) integrating local LAI values into WRF and designing comparative experiments; (iii) introducing the 3DVAR data assimilation method and designing comparative experiments with or without assimilation; (iv) based on the simulation results before and after optimization, calculating the wind power density and evaluating the impact of the multi-source optimization framework on the assessment of wind energy resources.

The scientific contribution of this study lies in constructing a multi-source collaborative optimization framework that integrates the seasonal selection of PBL parameterization schemes, local adjustment of LAI, and 3DVAR data assimilation based on the WRF model. It systematically evaluates the synergy effect of the three. It reveals the advantages of the YSU scheme, with unstable stratification in the spring the high layer in the winter, as well as the applicability of the local scheme near the ground in winter. It quantifies the cumulative improvement effect of local LAI adjustment and the introduction of data assimilation on the simulation of the near-ground wind field in forest areas and verifies the practical value of the collaborative framework. On this basis, this study further analyzes the impact of the multi-source optimization scheme on the assessment of wind energy resources and verifies the superiority of the collaborative optimization framework in wind field simulation in forest areas. The research results can provide scientific basis for wind farm site selection, turbine selection, and forest engineering design in the northeastern forest area, and also provide methodological references for wind field simulation optimization in other regions.

2. Materials and Methods

2.1. Study Area and Data Sources

The Maoershan forest region (Figure 1), located in Shangzhi City, Heilongjiang Province (127.48–127.73° E, 45.23–45.48° N), serves as a typical representative of the mountainous forests in eastern Northeast China. The vegetation is dominated by deciduous broadleaf forests and mixed coniferous–broadleaf forests, with a canopy coverage of approximately 95% and an average tree height of 16 m. The terrain is characterized by low hills and hills, with an average elevation of about 300 m. The region has a temperate monsoon climate, featuring long, cold, and dry winters, hot and humid summers, dry and windy springs, and mild autumns, making it a representative area of its kind. Within the region lies the Heilongjiang Maoshan Forest Ecosystem National Field Scientific Observation and Research Station, equipped with a lidar-based multi-layer wind speed and direction observation system. This provides high-quality, multi-height measured data to validate WRF simulation results. Therefore, the Maoshan forest area was selected as the research subject. Foliation typically begins in May, and leaf fall occurs in early September.

Observed wind data were obtained from a WindMast WP350 Doppler LiDAR (Figure 2a), installed at 45.403825° N, 127.659458° E, within the Maoershan Observation Station at an elevation of 389 m. The lidar operates on the principle of coherent optical pulse Doppler shift detection with a sampling frequency of 10 Hz. It continuously measures wind speed and direction profiles between 20 m and 350 m above ground level, with a measurement range of 0–70 m/s and an accuracy better than 0.1 m/s. The system outputs 10-minute averages of 11 meteorological parameters, including wind speed, wind direction, temperature, and pressure. As for the height selection for comparative analysis between simulation results and measured data, 20 m can be used to evaluate the direct effect of surface roughness and forest on near-surface wind shear because it corresponds to the area near the top of the canopy. As the transition layer, 50 m is helpful to analyze the characteristics of the near-ground disturbance transfer to the upper layer. However, 100 m corresponds to the typical hub height range (80–120 m) of modern large wind turbines, which is the core concern layer of wind energy resource assessment. At the same time, these three heights happen to coincide with the standard output layer of the lidar measured data in this study, so that the accuracy and reliability of the model validation can be ensured. Therefore, three heights of 20 m, 50 m and 100 m were selected for analysis and verification in this study. This selection helps to systematically reveal the vertical evolution law of the forest wind field from the near-surface dominated by local turbulence to the high-level dominated by large-scale forcing, and also provides a basis for evaluating the applicability of a multi-source optimization scheme in different physical characteristic layers.

2.2. WRF Model Configuration

Simulations were performed using WRF version 4.3.1 with a four-level two-way nested domain configuration (d01–d04). The horizontal grid resolutions and grid dimensions for each nested domain were 9 km (201 × 201), 3 km (217 × 217), 1 km (217 × 217), and 333 m (217 × 217), respectively. The domain center was set at 45.40° N, 127.66° E (Figure 2b). Vertically, 49 levels were configured, with the lowest six levels located at heights of 8.1 m, 24.3 m, 44.65 m, 73.15 m, 114.0 m, and 171.41 m above ground level. The model top pressure was fixed at 5000 Pa, and the Lambert conformal projection was applied. The National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) Final (FNL) reanalysis dataset served as the initial and boundary conditions, with the first 12 h of simulation discarded as spin-up. Since the abundant season of China’s wind power generation is concentrated in spring and winter, this study selected December and May for simulation [56,57]. Topographic data were derived from the Shuttle Radar Topography Mission (SRTM) digital elevation model at 90 m resolution [58]. Customized high-resolution terrain data were integrated into WRF following the procedure outlined by previous researchers to enhance the model’s representation of local topography.

2.3. Methodology for Parameterization Scheme Configuration

Previous studies [59] have demonstrated that the applicability of planetary boundary layer (PBL) schemes is region-and season-dependent, with significant variations in performance under different meteorological and surface conditions. Therefore, three representative PBL schemes—YSU (Yonsei University), MYNN (Mellor-Yamada–Nakanishi-Niino), and MYJ—were selected to design three simulation cases, enabling the identification of the most suitable parameterization scheme combination for the study area. Except for the PBL scheme, all other parameterization options were kept identical across the three cases. As the surface-layer and PBL schemes are inherently coupled, no additional modifications were required. The parameterization configurations for the three cases are listed in

Table 1. Configuration of three PBL scheme simulation cases.

Scheme	Case 1	Case 2	Case 3
Planetary boundary layer scheme	YSU	MYNN	MYJ
Surface layer scheme	MM5	MYNN	Eta
Land surface scheme	Noah-MP
Microphysical scheme	WSM6
Cumulus parameterization scheme	Kain–Fritsch cumulus potential
Radiation Scheme	RRTMG shortwave and longwave
Leaf area index (LAI)	Default values

, and the simulation period covered 1–30 May 2021 and 1–30 December 2020.

2.4. Methodology for Forest Canopy Characteristic Parameters

The leaf area index (LAI), defined as the ratio of total one-sided leaf surface area to the corresponding ground area, is a crucial canopy structural parameter closely related to wind distribution in forested ecosystems. Its mathematical definition is expressed as

$$\mathrm{LAI}={\displaystyle \frac{{\mathrm{S}}_{\mathrm{leaf}}}{{\mathrm{S}}_{\mathrm{land}}}}$$

(1)

where ${\mathrm{S}}_{\mathrm{leaf}}$ is the total leaf surface area, and ${\mathrm{S}}_{\mathrm{land}}$ is the projected ground area.

In the WRF model, the default LAI values are derived from global optimized datasets, which lack local specificity and thus fail to fully represent the aerodynamic effects of vegetation on local wind fields. Currently, LAI data are obtained either through direct or indirect methods. Direct methods, including destructive sampling, litter collection, and inclined-point sampling, are accurate but labor-intensive, time-consuming, and unsuitable for long-term monitoring due to their destructive nature. Indirect methods, based on optical principles, utilize instruments such as LAI-2200, TRAC, and SunScan to estimate canopy properties efficiently.

Specifically, the default LAI values in WRF are stored in the VEGPARM.TBL file, typically provided as 12 monthly datasets. This study directly replaces the decadal average LAI values of the tree canopy growth period obtained from Maoershan measurements into the corresponding land use types in VEGPARM.TBL, with the roughness length $z_0$ adjusted synchronously in VEGPARM.TBL. WRF performs linear interpolation on the VEGPARM.TBL data using its internal time step, enabling localized correction of canopy dynamics. The planar displacement height is dynamically diagnosed by the Noah-MP land surface scheme based on canopy attribute LAI isosurface data, eliminating manual intervention.

2.5. Methodology for 3DVAR Data Assimilation on the Simulated Wind Field in Forest Areas

As a representative technique of multi-source statistical data assimilation, the three-dimensional variational (3DVAR) scheme enables the assimilation of conventional observations into numerical weather prediction models. Its fundamental principle is to utilize the assimilated observations at each analysis time to generate an improved initial condition field. The model then integrates this field forward to the next assimilation cycle, where the procedure is repeated until the end of the forecast period. This cyclic assimilation framework provides high computational efficiency and supports the integration of diverse observational data sources, making it well suited for operational forecasting applications.

The mathematical foundation of the 3DVAR scheme is the minimization of a cost function to obtain the optimal analysis field ${\mathrm{x}}_{\mathrm{a}}$. The cost function, also referred to as the objective function, is defined as the weighted sum of the squared deviations between the analysis field $\mathrm{x}$ and the background field ${\mathrm{x}}_{\mathrm{b}}$, and between the simulated observation $\mathrm{x}$ and the actual observation ${\mathrm{y}}_{\mathrm{o}}$, as expressed in Equation (2).

$$\mathrm{J}\left(\mathrm{x}\right)={\displaystyle \frac{1}{2}}{\left(\mathrm{x}-{\mathrm{x}}_{\mathrm{b}}\right)}^{\mathrm{T}}{\mathrm{B}}^{-1}\left(\mathrm{x}-{\mathrm{x}}_{\mathrm{b}}\right)+{\displaystyle \frac{1}{2}}{\left[{\mathrm{y}}_{\mathrm{o}}-\mathrm{H}\left(\mathrm{x}\right)\right]}^{\mathrm{T}}{\mathrm{R}}^{-1}\left[{\mathrm{y}}_{\mathrm{o}}-\mathrm{H}\left(\mathrm{x}\right)\right]$$

(2)

where B denotes the background error covariance matrix, R is the observation error covariance matrix (including both instrumental and representativeness errors), and H represents the observation operator that projects model-space variables into the observation space. The term ${\mathrm{x}-\mathrm{x}}_{\mathrm{b}}$ is the background increment, and $({\mathrm{y}}_{\mathsf{\sigma }}-\mathrm{H}(\mathrm{x}\left)\right)({\mathrm{y}}_{\mathrm{o}}-\mathrm{H}(\mathrm{x}\left)\right)$ is the observation increment.

In this study, the WRF-3DVAR module was driven by the NCEP ADP global upper-air and surface observation datasets, which include pressure, geopotential height, temperature, dew point temperature, wind direction, and wind speed. In complex terrain forest areas, the efficiency of the 3DVAR system mainly depends on the representative error control of the observation operator (H) and the structural characteristics of the background error covariance matrix B. Due to the fact that the observation height of the LiDAR does not completely overlap with the non-uniform Eta coordinate layers of the WRF model, this study has constructed an observation operator H based on physical constraints. This operator abandons the traditional spatial linear interpolation and instead combines the leaf area index, displacement height and roughness length determined in Section 2.4 to perform a dynamic reconstruction of the logarithmic–linear wind speed profile in the near-surface layer. This improvement enables the observation operator to explicitly express the momentum sink effect of the forest canopy, effectively mapping the vertical wind field characteristics at the sub-grid scale to the model grid points, thereby improving the assimilation accuracy of the near-ground wind speed.

In addition, the 3DVAR process introduces geostrophic balance and static balance constraints and couples the vertical wind field increments with the temperature field and pressure field through multivariate analysis. In the undulating terrain of the Northeast region, this collaborative correction mechanism effectively suppresses the non-physical gravity wave noise in the initial assimilation stage and ensures the dynamical consistency in the complex stratified environment. The background error covariance matrix B was constructed using the NMC method. However, the experimental results show that the static B matrix has limitations in handling the seasonal surface forcing changes in the Northeast forest area. Therefore, this study constructed a seasonal-based background error covariance matrix for forest areas. Based on the historical simulation fields of spring and winter, the NMC method was used to independently calculate the B matrices for the two seasons in order to reflect the flow-dependent error characteristics of the atmospheric boundary layer structure in different seasons, replacing the default single B matrix of WRF. The WRFDA gen_be tool was run for spring and winter respectively, generating analysis fields by minimizing the objective function. The controlled variables employed the CV5 scheme, comprising the stream function (psi), unsteady velocity potential (chi_u), unsteady temperature (t_u), and relative humidity (rh). This scheme incorporates geostrophic and baroclinic equilibrium constraints, enabling the establishment of dynamic coupling between mass fields and wind fields through regression coefficients. Two experimental configurations were designed: Case 4, a control experiment without data assimilation, and Case 5, an experiment with assimilation. Simulations were conducted for the winter period (17–19 December 2020) and the summer period (16–18 May 2021).

2.6. Wind Energy Assessment Method

The simulation periods were set for the winter of December 2020 and the vigorous vegetation growth seasons of May 2021. A comprehensive assessment was conducted using three indicators: wind power density, annual energy production (AEP), and capacity factor (CF). Given Heilongjiang Province’s winter conditions, characterized by extreme sub-zero temperatures (potentially below −30 °C) and high-pressure systems, air density significantly exceeds the standard atmospheric value of 1.225 kg/m³. Consequently, this study employs dynamic air density, calculated using instantaneous pressure (P) and temperature (T) data from the WRF model. The methodology is detailed below.

$$\mathrm{P}\left(\mathrm{z},\mathrm{t}\right)={\displaystyle \frac{1}{2}}{\mathsf{\rho }(\mathrm{z},\mathrm{t})\mathrm{U}\left(\mathrm{z},\mathrm{t}\right)}^3$$

(3)

$$\mathsf{\rho }(\mathrm{z},\mathrm{t})=\mathrm{P}/(\mathrm{R}\cdot \mathrm{T})$$

(4)

where ρ denotes the air density (kg/m³), typically assigned a fixed value of 1.225 kg/m³; U represents the wind speed (m/s); and P(z,t) is the wind power density (W/m²) at height z and time t; $\mathrm{R}$ is the gas constant of dry air, which is 287 $\mathrm{J}/(\mathrm{k}\mathrm{g}\cdot \mathrm{K})$. P and T are the pressure (Pa) and temperature (K) at time t, respectively.

$$\mathrm{A}\mathrm{E}\mathrm{P}=8760{\int }_{{\mathrm{v}}_{\mathrm{i}\mathrm{n}}}^{{\mathrm{v}}_{\mathrm{o}\mathrm{u}\mathrm{t}}}{\displaystyle \frac{\mathrm{k}}{\mathrm{c}}}{\left({\displaystyle \frac{\mathrm{v}}{\mathrm{c}}}\right)}^{\mathrm{k}-1}\exp \left[-{\left({\displaystyle \frac{\mathrm{v}}{\mathrm{c}}}\right)}^{\mathrm{k}}\right]\cdot \mathrm{P}(\mathrm{v})\mathrm{d}\mathrm{v}$$

(5)

where ${\mathrm{v}}_{\mathrm{i}\mathrm{n}}$ and ${\mathrm{v}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$ represent the cutting-in and cutting-out wind speeds, respectively, with values of 3 m/s and 25 m/s; $\mathrm{k}$ and $\mathrm{c}$ are the shape parameter (m/s) and scale parameter (m/s) of the Weibull distribution; $\mathrm{P}\left(\mathrm{v}\right)$ is the wind turbine power curve, and the 2.5 MW model of the low-temperature type Goldwind Technology GW121/2500 (Xinjiang Goldwind Science & Technology Co., Ltd, Urumqi, China) is selected as the calculation basis.

$$\mathrm{C}\mathrm{F}={\displaystyle \frac{\mathrm{A}\mathrm{E}\mathrm{P}}{\mathrm{P}\times 8760}}\times 100\%$$

(6)

where $\mathrm{P}$ represents the rated power of the fan, which is set at 2.5 MWH.

To quantify the reliability of the WPD estimation, the error propagation method is introduced to conduct uncertainty analysis on the wind power density. The relative uncertainty calculation formula is as follows.

$${\mathrm{u}}_{\mathrm{r}\mathrm{e}\mathrm{l}}={\displaystyle \frac{\mathrm{P}\times \sqrt{{\left(3\frac{\mathsf{\Delta }\mathrm{v}}{\overline{\mathrm{v}}}\right)}^2+{\left(\frac{\mathsf{\Delta }\mathsf{\rho }}{\overline{\mathsf{\rho }}}\right)}^2}}{\overline{\mathrm{P}}}}\times 100\%$$

(7)

where $\mathrm{P}$ represents wind power density; $\mathsf{\Delta }\mathrm{v}$ is the deviation of average wind speed; $\overline{\mathrm{v}}$ is the average wind speed; $\mathsf{\Delta }\mathsf{\rho }$ is the deviation of air density; and $\overline{\mathsf{\rho }}$ is the average air density.

3. Results and Discussion

The flow chart of the multi-source optimization methods involved in this chapter is shown in Figure 3. In the Figure 3, the upward arrow indicates the degree of improvement in the MRE error indicator value. In order to verify the core assumptions and answer the research questions raised at the beginning of this paper, this chapter carries out the following four aspects of research content. Specifically, a variety of planetary boundary layer schemes PBL were selected for comparative analysis of simulation effects, and then the canopy characteristic parameter LAI was dynamically modified locally, and then 3DVAR assimilation data was introduced. Finally, the wind field simulation effect improvement analysis of multi-source optimization method was carried out from the perspective of wind energy assessment.

To accurately evaluate the simulation accuracy of wind speed and wind direction, six statistical indicators were employed: mean relative error (MRE), mean error (ME), root mean square error (RMSE), mean absolute error (MAE), Normalized Mean Bias (NMB) and correlation coefficient (R). Their mathematical formulations are expressed as follows.

$$\mathrm{MRE}={\displaystyle \frac{\mathrm{ME}}{\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{f}}_{\mathrm{o}}}},$$

(8)

$$\mathrm{ME}={\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{f}}_{\mathrm{i}}-{\mathrm{f}}_{\mathrm{o}}\right),$$

(9)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{f}}_{\mathrm{i}}-{\mathrm{f}}_{\mathrm{o}}\right)}^2},$$

(10)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{f}}_{\mathrm{i}}-{\mathrm{f}}_{\mathrm{o}}\right|$$

(11)

$$\mathrm{N}\mathrm{M}\mathrm{B}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{f}}_{\mathrm{i}}-{\mathrm{f}}_{\mathrm{o}}\right)}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{f}}_{\mathrm{o}}}}\times 100\%$$

(12)

$$\mathrm{R}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{f}}_{\mathrm{o}}-{\overline{\mathrm{f}}}_{\mathrm{o}}\right)\left({\mathrm{f}}_{\mathrm{i}}-{\overline{\mathrm{f}}}_{\mathrm{i}}\right)}{{\left[\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{f}}_{\mathrm{o}}-{\overline{\mathrm{f}}}_{\mathrm{o}}\right)}^2\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{f}}_{\mathrm{i}}-{\overline{\mathrm{f}}}_{\mathrm{i}}\right)}^2\right]}^{1/2}}},$$

(13)

where i represents the time index, n denotes the total number of samples, $f_i$ indicates the simulated value at time i, and $fo$ denotes the corresponding observed value. The ${\overline{f}}_i$ and ${\overline{f}}_o$ represent the mean values of the simulated and observed datasets, respectively.

When calculating the error indicators for wind direction, both the simulated wind direction values and the measured wind direction values are considered. At this time, since wind direction is a cyclic variable that varies within the range of 0° to 360°, the traditional linear calculation will produce incorrect jumps at the 0°/360° boundaries. Therefore, the values of $f_i-f_o$ in Formulas (9)–(12) need to be cyclically corrected to ensure that the error is within the range of [−180°, 180°]. The specific details are as shown in Formula (14).

$${(f_i-f_o)}^{\prime }=\{\begin{array}{c}{f}_i-f_o ,|f_i-f_o|\le 180^{°}\\ f_i-f_o-360^{°},f_i-f_o>180^{°}\\ f_i-f_o+360^{°},f_i-f_o<-180^{°}\end{array}$$

(14)

3.1. Results of Parameterization Scheme Configuration

The simulation results are analyzed.

Table 2. Error metrics between simulated and observed wind speeds for different PBL schemes in May and December.

Time	Height	Index	Case 1	Case 2	Case 3
May	20 m	RMSE (m/s)	2.11	2.29	2.27
		ME (m/s)	0.82	0.91	0.94
		MAE (m/s)	1.69	1.81	1.83
		NMB (%)	40.00	44.40	45.90
		R	0.51 *	0.49 *	0.49 *
	50 m	RMSE (m/s)	1.68	1.69	1.73
		ME (m/s)	0.58	0.67	0.90
		MAE (m/s)	1.34	1.34	1.39
		NMB (%)	21.20	24.50	33.00
		R	0.53 *	0.52 *	0.51 *
	100 m	RMSE (m/s)	1.36	1.83	1.91
		ME (m/s)	0.88	0.89	1.05
		MAE (m/s)	1.09	1.45	1.51
		NMB (%)	27.20	27.60	32.50
		R	0.57 *	0.55 *	0.57 *
December	20 m	RMSE (m/s)	2.12	2.03	2.00
		ME (m/s)	1.16	0.91	0.92
		MAE (m/s)	1.69	1.71	1.62
		NMB (%)	33.60	31.80	32.20
		R	0.50 *	0.48 *	0.48 *
	50 m	RMSE (m/s)	2.11	2.01	2.02
		ME (m/s)	0.90	0.85	0.84
		MAE (m/s)	1.69	1.68	1.71
		NMB (%)	26.50	32.00	34.10
		R	0.52 *	0.51 *	0.50 *
	100 m	RMSE (m/s)	1.95	2.32	2.47
		ME (m/s)	1.01	1.26	1.39
		MAE (m/s)	1.53	1.83	1.99
		NMB (%)	19.40	26.90	29.70
		R	0.55 *	0.53 *	0.54 *

* indicates that the significance level of the correlation coefficient reaches 0.05.

shows the error index and significance test results of simulated wind speed and measured wind speed at different heights in May and December under different planetary boundary layer (PBL) parameterization schemes. To objectively assess the statistical significance of the differences among the various schemes, this study employed the paired sample Wilcoxon signed-rank test and conducted a Fisher Z transformation comparison test on the correlation coefficients (R) among the schemes to examine the potential systematic deviation between the simulated values and the observed values at a 0.05 significance level.

As shown in Table 2, the root mean square error of all experiments reached its peak at a height of 20 m (RMSE > 2.1 m/s), mainly due to the strong aerodynamic resistance within the forest canopy–atmosphere interface and the sub-grid scale terrain turbulence. As the altitude increased, the correlation coefficient of Case 1 showed a gradually increasing trend, indicating that the WRF model has an outstanding ability to capture the mesoscale weather forcing and wind speed dynamics in the upper part of the planetary boundary layer (PBL). At 100 m height, the absolute error increased, which was due to the increase in wind speed with altitude, and the R increased, which was due to the fact that the turbulence in the wind field is more stable than that near the ground. Overall, Case 1 consistently demonstrated the most robust prediction performance in all vertical layers. In May, all the indicators of Case 1 remained at the leading position. At the 20 m and 50 m heights, the correlation coefficient of the YSU scheme was significantly higher than that of the MYJ and MYNN schemes, indicating that YSU has superior ability in simulating the wind speed trend under the spring unstable stratification. In December, Case 1 showed slightly higher RMSE and NMB values at the 20 m near-ground height. This was attributed to the advantages of Case 2 and Case 3 in using the local TKE closure scheme to simulate the near-ground winter stable stratification, while YSU was limited by the physical constraints of excessive mixing. However, at an altitude of 100 m, the R value of YSU (0.55) was still higher than that of MYJ and MYNN, but there was no significant difference.

Overall, in spring, YSU demonstrated significant advantages at all heights. In winter, although YSU could also simulate wind speed relatively well near the ground, its performance was slightly inferior to the local closed-loop scheme. However, as the height increased, the simulation ability of YSU narrowed the gap with other schemes even more. Moreover, this indicates to some extent that YSU has excellent simulation performance in both spring and winter seasons, and maintains stable performance in simulating mountain forest wind fields with strong dynamic shear.

It is worth noting that all three schemes showed positive mean error (ME) values (ranging from 0.58 to 1.05 m/s), indicating a systematic overestimation of wind speed. This suggests that even with the YSU and MM5 schemes, WRF still exhibits systematic overestimation when dealing with high roughness surface layers. Part of the reason is that the current schemes cannot capture the dynamic influence of the near-ground vegetation roughness on the wind field, and the model’s portrayal of near-ground momentum decay and energy dissipation is insufficient. Therefore, it is necessary to introduce vegetation feature descriptions in the WRF model to improve the accuracy of wind field simulation in the forest areas of Northeast China.

The wind direction simulation errors at different heights are summarized in

Table 3. Comparison of error parameters between simulated and observed wind directions at different altitudes in May and December.

Time	Height	Index	Case 1	Case 2	Case 3
May	20 m	ME (°)	−6.11	−16.11	0.94
		RMSE (°)	107.57	114.85	1.78
		MAE (°)	59.63	93.02	76.23
	50 m	ME (°)	−21.24	−32.06	−15.78
		RMSE (°)	87.87	92.02	91.80
		MAE (°)	67.26	69.01	70.33
	100 m	ME (°)	−2.14	−3.07	0.45
		RMSE (°)	89.93	106.00	107.50
		MAE (°)	49.87	55.22	56.32
December	20 m	ME (°)	−10.12	−8.73	−11.39
		RMSE (°)	87.02	86.25	89.03
		MAE (°)	57.35	54.47	52.59
	50 m	ME (°)	−13.23	−12.67	−13.01
		RMSE (°)	76.77	76.31	77.27
		MAE (°)	51.38	50.19	52.40
	100 m	ME (°)	−14.18	−13.87	−12.27
		RMSE (°)	43.23	44.39	49.11
		MAE (°)	20.72	21.45	20.33

.

As shown in Table 3, under weak wind conditions, wind direction in forested areas exhibits strong randomness and localized characteristics. The Northeast Forest Region demonstrates significant thermal inhomogeneity in May, with complex canopy structures frequently causing weak instability or neutral stratification in the near-surface layer. The YSU scheme achieves optimal RMSE and MAE values across three altitudes, with most ME values approaching optimal levels, demonstrating strong capability in capturing wind direction variations under spring instability stratification. This stems from YSU’s non-local approach, which fully accounts for full-layer mixing effects induced by large vortex thermal driving, while MYNN and MYJ schemes as local (Local) approaches primarily focus on gradient diffusion between adjacent layers. The MM5 scheme, based on the Moinin–Obhoff similarity theory and incorporating stability correction with segmented function optimization for stability parameters, exhibits greater robustness in handling friction velocity. Comparative results from December reveal MYNN’s slight advantage in lower layers and YSU’s minimum RMSE in upper layers. This reflects that local closure schemes better maintain near-surface structures under winter stability stratification, while non-local approaches can also simulate fundamental wind direction characteristics at higher altitudes. The December comparative analysis of various schemes demonstrates MYNN’s marginal advantage in lower layers, while YSU achieves the lowest RMSE in upper layers. This indicates that local closure schemes under stable winter stratification better preserve near-surface structures, though non-local schemes can also accurately simulate fundamental wind direction characteristics in upper layers.

In summary, the integrated wind speed and direction simulation results for both summer and winter demonstrate that the YSU scheme achieves optimal overall performance in spring. Although its near-surface wind direction in winter is slightly inferior to MYJ, its overall error level remains comparable to MYJ and MYNN, indicating low sensitivity. Therefore, this section answers the research question Q1 by comparing the simulation errors of different PBL schemes in May and December and at different heights: The YSU scheme has the best performance in unstable stratification in May and in high layers in winter, and its RMSE is reduced by up to 25.7% compared with other schemes. The MYJ scheme has a slight advantage in the surface layer only in December (the RMSE difference is within 6%, and the R value is not much different), which reveals that PBL scheme shows inter-seasonal applicability in the simulation of a mountain forest wind field. The selection of the PBL parameter scheme is fully justified. However, the single mesoscale parameterization scheme still exhibits significant limitations in fine-scale forest canopy simulation, such as overestimating wind speeds, which requires further optimization.

Based on the above analysis, the main conclusions of this study can provide the following guidance for other forest areas. In the unstable stratification of spring, priority should be given to non-local solutions (such as YSU); in the stable stratification of winter, if the research focus is on the near-surface wind field (such as in forest ecological studies), local solutions (such as MYJ) can be considered; if the focus is on the upper-level wind field (such as at the height of the wind turbine hub), YSU remains the more reliable choice.

3.2. Results of Forest Canopy Characteristic Parameters LAI

The dominant tree species in the Maoershan region include Xing ‘an deciduous pine, Korean pine, red pine, birch, Mongolian oak, poplar and water elm. According to the MODIS land surface classification, these correspond to Type 4 (deciduous broadleaf forest) and Type 5 (mixed forest). Considering the local forest composition, simulations were conducted during May–June, the active growing season characterized by rapid canopy development. Based on field LAI measurements obtained using the LAI-2200 system [60], localized LAI corrections were applied to the study area. The crown height was 16 m, and the minimum and maximum values of roughness length were adjusted to 1.3 m and 1.9 m for deciduous broadleaf forest, and 1.4 m and 1.9 m for mixed forest. The default LAI values were adjusted every ten days to capture the dynamic canopy development during the leaf-expansion period, as summarized in

Table 4. Configuration of the LAI cases.

Canopy Feature Parameters	Time	Case 1		Case 4
		Default		Local Fix
		Deciduous Broadleaf Forest	Mixed Forest	Deciduous Broadleaf Forest	Mixed Forest
Leaf area index (LAI)	May	3.0	3.5	2.0, 3.8, 4.5	3.0, 4.1, 4.5
	June	4.7	4.3	5.0, 5.6, 6.0	5.0, 5.7, 6.0

.

Figure 4 presents a comparative illustration of the discrepancies between simulated wind speed profiles before and after LAI correction and corresponding observational data.

Table 5. Comparison of simulated wind speed and direction values with measured data before and after LAI correction at different heights (MRE).

Index	Case	20 m	50 m	100 m
MRE (wind speed)	Case 1	0.403	0.189	0.249
	Case 4	0.262	0.117	0.236
MRE (wind direction)	Case 1	−0.078	−0.136	−0.011
	Case 4	−0.055	−0.132	0.010

quantifies the results, providing an error analysis of simulated versus observed wind speeds and wind directions at various heights prior to and following the LAI localization correction.

Figure 4 results indicate that the default LAI (Case 1) significantly overestimates simulated wind speeds at 20 m height compared to observations, reflecting an underestimation of canopy density in northeastern forest regions. Following localized correction (Case 4), the wind profile gradient markedly diminishes, demonstrating that enhanced canopy resistance produces a more realistic deceleration effect on the lower atmosphere.

Furthermore, the quantitative analysis in Table 5 shows that the model has reduced the average relative error (MRE) of wind speeds at 20 m, 50 m, and 100 m by 0.141, 0.072, and 0.013, respectively. The accuracy at 20 m and 50 m has increased by approximately 35%. This indicates that the model has a clear and consistent physical response to changes in the LAI parameter. Additionally, the corrected LAI parameter can more accurately reflect the true canopy structure characteristics of the vegetation, thereby realistically reproducing the influence of canopy resistance on turbulence during the simulation process. The increase in LAI directly enhances the physical obstructive area of the canopy to the airflow, resulting in a significant increase in the dynamic roughness. As the localized LAI depicts a more dense forest canopy structure, the zero plane moves upward. According to the logarithmic wind profile theory, after the localized LAI corrects the over-sparse deviation that is commonly present in the default dataset in the northeastern forest area, the increase in roughness and the displacement of the zero plane jointly leads to an enhancement of the momentum sink at a height of 20 m, so the improvement value of MRE at 20 m is as high as 0.141. This correction enables the simulation process to more realistically reproduce the reduction effect of the canopy interior and edge on turbulence. In addition, the influence of LAI correction on wind speed simulation is most significant below the 20 m height. This is because LAI is closely related to the density of the vegetation and the configuration of the canopy; above the canopy height, LAI cannot directly exert its effect but indirectly affects the high-altitude wind field situation through the vertical transfer of turbulent shear stress, and the energy gradually dissipates as the height increases.

All three simulation cases confirmed the effectiveness of local LAI correction, demonstrating that incorporating site-specific canopy data significantly enhances wind speed simulation accuracy. In contrast, the improvement in wind direction accuracy was relatively modest: the Mean Root Square Error (MRE) at 20 m decreased by only 0.023, representing a 28% accuracy gain; however, the accuracy changes at 50 m and 100 m were negligible. The reason for this is that the turbulence intensity in the forest area is relatively high, and the turbulence structure is complex. The thermal effects in the forest area (such as surface heating and cooling) influence the formation of the local wind field, and there are systematic deviations in the large-scale background field.

In summary, this section answers the research question Q2 by comparing the simulation results before and after LAI correction: The correction of localLAIparameters effectively made up for the deficiency of the default data set in describing the characteristics of forest areas, reducing the wind speed error by 35% and improving the wind direction accuracy by 28%. The optimization effect was attenuated with the height and enhanced with the increase in LAI, which was due to the fact that LAI was a characteristic parameter describing the influence of the near-surface canopy on the wind field. After the local dynamic correction, the improvement effect is particularly prominent near the ground.

This indicates that the local adjustment of LAI should be included as a standard configuration in the WRF simulation of forest areas, especially for the simulations in the summer (when LAI is high). In the next stage, based on the corrected vegetation canopy characteristic parameters in this section, 3DVAR three-dimensional variational assimilation technology will be introduced. By integrating multi-source conventional meteorological observation data, the system deviation will be improved, and the deep optimization of the temporal and spatial evolution process of the forest area wind field will be achieved.

3.3. Results of 3DVAR Data Assimilation

The trial period was selected from 00:00 UTC on 1 May 2021 to 00:00 UTC on 1 June 2021, and from 00:00 UTC on 1 December 2020 to 00:00 UTC on 1 January 2021. Cold-start 24 h forecasts were conducted daily based on initial fields at 00:00 and 12:00. Using two sets of forecast fields simulated by the WRF model at the same time points, with forecast lead times of 12 h and 24 h, respectively, as sample data, the background error covariance B was subsequently calculated. A periodic assimilation strategy with six-hour intervals was employed, with the assimilation time window set to t ± 3 h. Observation data were read at 00:00, 06:00, 12, and 18 h (UTC). Two experimental configurations were designed: Case 4, a control experiment without data assimilation, and Case 5, an experiment with assimilation. Simulations were conducted for the winter period (17–19 December 2020) and the summer period (16–18 May 2021).

The results were compared to evaluate the effect of data assimilation on improving the accuracy of the simulated wind field. Figure 5 and Figure 6 illustrate the simulated wind speed results for the two seasons.

As shown in Figure 5, both Case 4 and Case 5 reproduced the temporal evolution of the observed near-surface wind speed at three height levels. As height increased to 50 m, Case 5 exhibited significantly improved agreement with the measured data compared with Case 4, capturing both the magnitude and fluctuation characteristics of the observed wind speed. At 100 m height, the simulated curve nearly coincided with the observed curve, and the bias between Case 4 and the measurements was substantially reduced. The wind speed in the forested region generally increased between 00:00 and 12:00 (local time), reaching its daily maximum around noon.

By comparing assimilation with non-assimilation schemes in December versus May, it is evident that data assimilation significantly improves WRF’s temporal and amplitude corrections for wind speed simulations during the December period. The hourly comparison in Figure 6 demonstrates that Assimilation Scenario 5 accurately captures extreme events’ timing and intensity at all three observation heights during the trough at 07:00 on 18 December and the peak at 13:00. In contrast, Non-Assimilation Scenario 4 exhibits significant temporal lag and amplitude deviation at 50 m, showing persistently low wind speeds starting at 12:00 on 18 December, maximum errors by 17:00, and peak errors exceeding 3 h. Combined with field observation records of sudden rainfall events, this indicates that short-term intense convective processes triggered by precipitation caused abrupt changes in near-surface wind fields. These effects were inadequately captured in the non-assimilated model, leading to increased deviations in Case 4. Therefore, meteorological observation assimilation data plays a corrective role in WRF model calculations.

Figure 7 shows the time evolution process of wind direction at 20 m, 50 m, and 100 meters during the observation and simulation data in May. The observation data indicates that the wind direction fluctuations throughout the simulation period exhibit significant non-stationary characteristics. Weather analysis suggests that these oscillations are mainly driven by convective activities accompanying precipitation events. Overall, the observation agreement of Case 5 is significantly better than that of Case 4. From 15:00 on May 16th to 00:00 on May 17th, a transition in the key wind direction from north wind to east wind was observed. However, Case 4 failed to timely capture this transition and exhibited significant phase lag and systematic undemrestimation at other times. In contrast, Case 5 accurately captured the physical characteristics and intensity changes in this directional shear throughout the time series and at various heights. This improvement fully demonstrates that by optimizing the 3DVAR assimilation method of initial conditions, the model can significantly improve its accuracy in representing large-scale circulation dynamics.

Furthermore, influenced by the underlying heterogeneous forest, the wind fields at 20 m and 50 m exhibit high-frequency fluctuation characteristics. Case 5 shows stronger sensitivity to these small-scale disturbances. Although both schemes tend to underestimate the intensity of the wind direction transition, the residuals of Case 5 are always lower than those of Case 4. For example, a sudden drop in directional wind speed occurred suddenly in the early morning of 17 May. In conclusion, these results verify that the 3DVAR assimilation method not only constrains the background field but also constrains the stable boundary layer wind field structure through dynamic balance constraints.

Figure 8 further demonstrates that at 20 m, 50 m, and 100 m, the assimilated simulations (Case 5) yielded wind direction results consistently closer to the observations, with enhanced temporal stability. The results show that during this period, the wind direction in the forest area mainly concentrated between 250° and 350°, exhibiting a stable west–northwest prevailing wind characteristic. Compared with Case 4 (without assimilation scheme), Case 5 (3DVAR assimilation) has a more advantageous phase control in the reporting stage and the period of intense fluctuations. For example, around 15:00 on 17 December, the measured wind direction showed a significant deflection towards the northwest. Case 5 was able to more quickly fit this evolution trend and reduce the accumulation of initial field deviations.

The quantitative error analysis of wind speed and direction simulations at various altitudes in May 2021 and December 2020, after incorporating the WRF-3DVAR assimilation module, is presented in

Table 6. Error comparison between simulated and measured values of wind speed and wind direction before and after assimilation at different heights in May and December.

Time	Index		Case	20 m	50 m	100 m
May	Wind speed	R	Case4	0.609	0.635	0.682
			Case5	0.735	0.772	0.803
		RMSE (m/s)	Case4	1.102	4.299	1.392
			Case5	1.128	1.508	1.376
		MAE (m/s)	Case4	0.877	3.827	1.086
			Case5	0.907	1.268	1.097
	Wind direction	R	Case4	0.461	0.547	0.468
			Case5	0.690	0.708	0.713
		RMSE (°)	Case4	76.749	81.594	59.871
			Case5	29.663	28.893	25.106
		MAE (°)	Case4	54.084	60.696	45.102
			Case5	24.928	22.710	21.105
December	Wind speed	R	Case4	0.706	0.755	0.764
			Case5	0.811	0.852	0.869
		RMSE (m/s)	Case4	1.262	2.090	1.999
			Case5	0.668	1.761	1.727
		MAE (m/s)	Case4	1.014	1.781	1.524
			Case5	0.505	1.597	1.312
	Wind direction	R	Case4	0.603	0.637	0.645
			Case5	0.758	0.783	0.813
		RMSE (°)	Case4	27.020	24.740	23.463
			Case5	8.287	8.148	8.064
		MAE (°)	Case4	22.332	19.778	19.302
			Case5	6.840	6.348	6.328

.

Table 6 presents a quantitative comparison of the simulation errors of wind speed and wind direction at different heights between May 2021 and December 2020. In May, the correlation coefficients of wind speed at all heights exceeded 0.7. The assimilation significantly improved the accuracy of wind direction simulation within the boundary layer, especially in high-altitude areas, where the effect was more pronounced. This improvement stemmed from the fact that the upper-level flow in the boundary layer was mainly dominated by pressure gradients and Coriolis forces, while surface friction and thermal disturbances were relatively weak. Therefore, the WRF model’s wind direction calculation based on statics was more accurate. In December, the assimilation raised the correlation coefficient of wind speed at 20 m to 0.8 and the correlation coefficient of wind direction to 0.75. The improvement in winter was weaker than that in summer, and this seasonal difference reflected the influence of the cold climate characteristics of the region: the strong temperature inversion and stable boundary layer in winter suppressed turbulence and enhanced local terrain and thermal forcing effects. Overall, the assimilation effects in both May and December were favorable. The reason is that the boundary layer is more active in the growth season of May, and the vertical mixing caused by the forest canopy is significantly enhanced. At this time, the B matrix shows a wider vertical correlation scale, and the observation increment can effectively propagate upward from the canopy top to the upper part of the boundary layer. In the leaf-fall period of December, affected by strong radiation cooling, stable inversion layers frequently occur in the near-surface layer. The increase in stratification stability leads to a contraction of the vertical characteristic scale of the B matrix, significantly improving the physical adaptability of the assimilation system on the complex forest surface. Additionally, snow cover would change the surface albedo and roughness, and the current assimilation configuration did not explicitly consider the land surface with snow cover.

Compared with the studies by Cao et al. [3] which only focused on the localization of LAI, the various evaluation indicators for each height in this study have shown significant improvements. This verifies the effectiveness of the multi-source optimization framework that combines PBL optimization, LAI localization, and data assimilation, and can effectively enhance the simulation accuracy. The simulation results of this model provide a more reliable basis for high-resolution wind resource assessment in forest areas. Moreover, uncertainties exist within the analysis domain itself and the model configuration. For instance, snow cover alters surface albedo and roughness, yet the current assimilation setup does not explicitly account for snow-covered terrain. Consequently, these uncertainties propagate into subsequent wind speed and direction forecasts. Future work should seek methods to improve this process, thereby better quantifying and constraining such uncertainty propagation.

Overall, this section answers the research question Q3 by comparing the simulation results of assimilated and non-assimilated schemes: The 3DVAR assimilation significantly reduces the uncertainty of the low-level wind field, and shows better performance in the inertial subregion with more uniform physical conditions, which makes the correlation coefficient of wind speed up to 0.869 and the wind direction speed up to 0.813, confirming the effectiveness of constructing different B matrices for different seasons.

It is worth noting that the operational steps of the assimilation scheme do not depend on a specific location. As long as the study area has basic meteorological observation data, it can be replicated and implemented. This provides a direct applicable technical route for other forest areas. Moreover, for regions with obvious seasonal changes, using a single B matrix will introduce additional errors. Therefore, when applying 3DVAR assimilation in other forest areas, other forest areas should independently construct different-season B matrices based on the historical simulation field to more accurately reflect the error characteristics of flow-dependent effects. Furthermore, the physical constraint observation operator constructed in this study significantly improves the assimilation accuracy of near-surface wind speed by explicitly expressing the momentum sink effect of the canopy. This improvement idea is not only applicable to forest areas but can also be extended to other underlying surfaces with obvious canopy effects (such as urban canopies, high-stature crop areas), providing a method reference for data assimilation of complex underlying surfaces.

3.4. Results of Wind Energy Assessment

Based on the optimized parameterization scheme (YSU), the locally corrected canopy parameter (LAI), and the 3DVAR-assimilated data, wind power density was calculated for different simulation periods. The conventional WRF configuration was denoted as TWRF, while the optimized WRF configuration incorporating parameterization optimization, LAI correction, and data assimilation was denoted as OWRF, and both were used for comparative wind energy evaluations.

Figure 9 presents the wind power density at different heights for May 2021. The results exhibit pronounced vertical spatial heterogeneity, providing essential data for analyzing the vertical gradient of wind energy resources in the forested region. At the 20 m height, the overall trend of wind power density simulated by the optimized OWRF scheme aligns more closely with the measured values compared with the traditional TWRF scheme. Between 20:00 on 17 May and 03:00 on the following day, the wind power density at the meteorological tower site was relatively high, reaching two distinct peaks at approximately 60 W/m².

The underlying reason is that at lower altitudes, the forest canopy coverage is high, and the terrain exhibits significant undulations, resulting in enhanced surface friction and stronger influence of local circulation, which in turn lead to lower wind power density near the surface. The OWRF scheme effectively accounts for the influence of near-surface canopy characteristics on the wind field, enabling the model to accurately reproduce the timing and magnitude of the observed wind power peaks. In contrast, the traditional WRF scheme fails to capture the physical features of wind field variability near the surface, resulting in weaker performance in reproducing the temporal evolution and peak characteristics of wind power density.

Similarly, analysis of the wind power density at 50 m and 100 m (Figure 9) indicates that the maximum values reach 79 W/m² and 290 W/m², respectively. Aside from these peaks, wind power density generally increases with height, indicating that the influence of terrain roughness and vegetation friction diminishes at higher elevations, leading to higher overall wind power density and richer wind resources. The modulating effect of terrain on the wind field exhibits a clear vertical attenuation, while atmospheric processes gradually dominate the wind field structure.

At both 50 m and 100 m heights, the two experiments reproduce the temporal variations in observed wind power density with good agreement in the overall trend; however, OWRF successfully captured two power density peaks appearing in the afternoon of 17 May, whereas TWRF produced only one peak and exhibited larger simulation errors. This improvement is attributed to the combined optimization of the YSU planetary boundary layer scheme and data assimilation, confirming that OWRF demonstrates superior overall performance.

Figure 10 shows that in December 2020, the spatial heterogeneity of wind power density with height remains significant. At 20 m height, wind power density ranges from 30 to 130 W/m², slightly higher than in May. For the extreme wind power values, the simulation error of OWRF is 10 W/m², compared with 72 W/m² for TWRF, while the overall RMSE of OWRF is 9 W/m² lower than that of TWRF. At 50–100 m heights, wind power density increases substantially, expanding across 300–450 W/m², indicating a much greater availability of wind energy than in May.

Overall, the OWRF simulation results agree more closely with the observations and exhibit larger wind power density maxima than in spring. This is primarily due to the strong northwesterly winds generated by the pressure gradient between the Siberian High and the North Pacific Low during winter. The sparser winter canopy also reduces surface roughness, weakening the terrain-induced dynamic forcing and thus increasing overall wind speeds and power density. As altitude increases, the resistance effects of vegetation and topography are further reduced, resulting in higher wind power density and more abundant wind resources.

To further explore the reasons for the differences in wind power generation density, the Weibull distribution function was adopted for fitting and parameter analysis. The results are shown in

Table 7. The Weibull fitting results of the simulation results before and after assimilation at different heights in May and December.

Time	Height (m)	Case	k	c
May	20	TWRF	1.510	2.921
		OWRF	1.506	2.694
		Measured	1.527	2.278
	50	TWRF	2.165	6.504
		OWRF	2.252	3.801
		Measured	2.445	3.066
	100	TWRF	2.593	5.867
		OWRF	1.803	4.031
		Measured	1.677	3.615
December	20	TWRF	3.888	4.461
		OWRF	3.560	4.071
		Measured	2.661	3.700
	50	TWRF	2.709	4.999
		OWRF	2.528	4.443
		Measured	1.624	3.205
	100	TWRF	3.770	5.851
		OWRF	2.902	5.320
		Measured	2.682	5.273

.

In spring, the c values of OWRF at various heights were closer to the observed values than those of TWRF. Taking 50 m as an example, the c value of TWRF was too high, resulting in a severe overestimation of its peak wind power density; while OWRF, through the YSU scheme to optimize the local turbulent mixing process and combined with LAI to accurately depict the drag effect of the forest canopy, reduced the c value error to 24%, and the corresponding wind power density was more in line with the observation. This indicates that the accurate simulation of the mean wind speed by the OWRF scheme is the basis for improving the accuracy of wind power density estimation. In winter, at a height of 20 m, the k value of OWRF was slightly higher than the observed value, indicating that the simulated wind speed distribution was more concentrated than the measured one. This may be related to the fact that the YSU scheme still has certain deviations in depicting turbulent mixing in the near-ground layer under stable stratification; at a height of 100 m, the c value of OWRF was almost the same as the observation, while TWRF was 11% higher. This shows that the optimized OWRF effectively constrained the large-scale circulation background and accurately reproduced the high-altitude wind speed characteristics under the strong northwest wind in winter, explaining the high agreement between OWRF and the observation at the high-altitude wind power density in Figure 10. Overall, the OWRF scheme enhances low-level simulation accuracy by optimizing near-surface physical process descriptions, while effectively suppressing error propagation to higher layers, thereby systematically improving the reliability of upper-level wind energy assessment.

In conclusion, the Weibull parameter analysis confirmed that the OWRF scheme improved the simulation ability of the mean wind speed and distribution characteristics through the YSU scheme to optimize the turbulent process, LAI to correct the canopy drag effect, and data assimilation to constrain the large-scale background field, effectively improving the accuracy of wind power density estimation. The accurate characterization of the scale parameter c is the core advantage of the optimization scheme.

The calculations for TWRF, OWRF and the AEP and CF of the observed values at a height of 100 m yield an annual actual power generation total of 5668.2 MWH, with a capacity factor of 25.88%. The AEP and CF of TWR are 11,660.8 MWH and 53.25% respectively, while those of OWRF are 6365.7 MWH and 29.07% respectively. It can be seen that the default TWRF model leads to a significant distortion in the assessment of wind energy potential, and the annual power generation assessment value is overestimated by 105.7% (confidence level = 53.25%). This deviation mainly stems from the neglect of canopy resistance in forest areas, especially being more pronounced during the growing season. In contrast, the OWRF model that integrates dynamic leaf area index (LAI) and three-dimensional variational (3DVAR) data can significantly improve the consistency between the assessment results and the observed data. The annual AEP error is significantly reduced to 12.3% (confidence level = 29.07%). Research on the Romanian Carpathians [17] also indicates that vegetation significantly alters the temporal and spatial distribution of wind energy potential. The dynamic changes in vegetation driven by land use changes may have a drastic impact on the wind energy potential of complex terrain. This also suggests that the assessment of wind energy in forest areas must take into account the seasonal dynamics and long-term changes in canopy characteristics. The significant improvement in December data (error of only 4.2%) indicates that OWRF can accurately capture the stable wind energy potential during the dormant season, providing a scientific basis for seasonal energy scheduling in high-latitude forest areas.

By conducting error propagation analysis to quantify the uncertainty of the WPD assessment, it can be calculated that the relative uncertainty of the optimized OWRF scheme at a height of 20 m in spring decreased from 46.80% to 14.06%, and as the height increased, the uncertainty decreased. This indicates that with the weakening of the influence of complex canopy turbulence, the model stability was enhanced. In winter, it decreased from 30.61% to 9.06%, effectively verifying that the OWRF scheme can effectively improve the reliability of wind energy assessment.

Based on the wind field simulation results after multi-source optimization, this section calculates the wind power density, Weibull parameters, etc., and fully answers the research question Q4: The multi-source optimization can comprehensively improve the key indicators of wind energy assessment, especially in May, the c value at the height of 50 m is optimized by 78.6% compared with the traditional WRF model, and the MRE at the height of 20 m is reduced by 35%, which is relatively limited. This may be because the wind field at this height is affected by multiple factors such as terrain turbulence, thermal drive and local circulation in addition to canopy drag. In general, the results successfully verify that the optimization method can improve the accuracy of wind energy resource assessment and has engineering value.

4. Conclusions

To address the accuracy limitations of traditional WRF models in simulating wind fields in forested areas, this study proposes a multi-source optimization method that integrates localized canopy parameters (LAI) with 3DVAR data assimilation. This approach significantly improves wind field reproduction capabilities in complex terrain forested areas. Furthermore, it reveals the spatiotemporal distribution patterns of wind energy resources, offering practical reference value. The main research conclusions are summarized in three key points.

(1) Through comparative analysis of various planetary boundary layer parameterization schemes, it was found that the YSU scheme significantly improves the accuracy of forest wind field simulation results compared to MYNN and MYJ schemes. Specifically, the non-local YSU scheme performs best in the spring (with unstable stratification), while the local MYJ scheme has a slight advantage near the ground in the winter (with stable stratification). This pattern stems from the physical nature of different closed schemes—the non-local scheme takes into account the full-layer mixing of thermal large-scale vortices through the anti-gradient term, making it more suitable for convective active unstable stratification; the local scheme assumes that turbulence is determined solely by local gradients and can better maintain the inversion structure under stable stratification.

(2) The forest canopy structure plays a crucial role in numerical simulations of near-surface wind fields in forested areas. Based on measured LAI-2200 data, after applying localized corrections to the leaf area index (LAI) data representing local forest canopy characteristics on a decadal scale, the average relative error of near-surface wind speed decreased by 35%, while wind direction simulation accuracy improved by approximately 28%, significantly enhancing the precision of simulation results. This indicates that the local adjustment of LAI can effectively reduce the systematic deviation in the simulation of near-surface wind speed, and the improvement effect is more obvious in summer (with high LAI). This finding validates the momentum absorption effect of the forest canopy on the wind field and its seasonal variation characteristics. The physical mechanism—the canopy affects the near-surface wind profile by increasing surface roughness and momentum sink—is universal. Therefore, any WRF simulation of forest areas should consider the seasonal dynamics of LAI instead of using fixed default values of the climatic state.

(3) By further integrating the WRF-3DVAR data assimilation method into the WRF model and comparing it with the traditional WRF model without assimilation data, the results showed that the WRF-3DVAR assimilation method demonstrated better fitting performance in wind energy assessment during both winter and summer. Specifically, the improvement in wind speed simulation accuracy is relatively limited in winter, while the enhancement is more significant in summer. The correlation coefficient R value for wind direction can be increased to the highest of 0.813, and the R value for wind speed can be increased to the highest of 0.869. The improvement is most significant in the surface layer, which is most directly affected by the canopy, and naturally decreases with the increase in height. In summer, when the canopy closure is high, the improvement is better than in winter, when the canopy is sparse. This vertical distribution feature just verifies that the multi-source optimization method in this study is successful in characterizing the forest canopy momentum absorption effect, and the model accurately captures the physical mechanism of the canopy as a momentum sink, so that the improvement effect is concentrated in the near stratum where the physical process is the most complex. In addition, the strategy of constructing the background error covariance matrix B by season is confirmed to be effective, and the simulation results in May and December are highly consistent with the observations, which provides a more scientific reference for the assessment of wind energy in forest areas. It provides a more scientific reference basis for actual wind energy assessment and development, offering insights for further optimization of forest area wind field numerical simulation research.

(4) The wind power density at 100 m altitude in the Maroershan can reach 475 W/m², 38% higher than that in summer, indicating that the forest area of Maoershan has stable and exploitable wind energy resources. The vertical wind energy stability also shows marked improvement, with resource value increasing with altitude. The area proves particularly valuable for wind energy development at elevations above 100 m.

Research findings demonstrate that the WRF optimization approach integrating local LAI correction with data assimilation significantly enhances the accuracy of near-surface wind field simulations in forested areas. This advancement provides scientific support for wind energy development, wind turbine deployment, and windbreak engineering in forest regions. It also improves the applicability of the WRF model under complex vegetation cover and non-flat terrain, offering valuable references for forestry meteorological research. This method enhances the applicability of the WRF model in complex vegetation coverage and non-flat terrain conditions, successfully forming a clear, replicable optimization method based on WRF simulations for detailed simulation of wind fields in forest areas. This method itself is not dependent on specific locations. For instance, the non-local YSU scheme performs best under unstable spring stratification, the local MYJ scheme has a slight advantage near the ground in stable winter stratification, but the simulation of the upper-level wind field still favors YSU; the local LAI correction should be included as a standard configuration for WRF simulations in forest areas, especially for simulations of high LAI in summer; and when conducting assimilation experiments, different seasons should be separately constructed B rather than using a single B matrix. The method of constructing an observation operator based on LAI and canopy parameters can be extended to other canopy substrates. These findings are not accidental results specific to a certain location. The research framework and physical laws can be generalized to wind field simulations and wind energy assessment studies in other forest areas. Future studies could further integrate seasonal variations and complex weather patterns to achieve more precise simulations of turbulent structures.

This study constructed a multi-source collaborative optimization framework integrating PBL scheme optimization, LAI local dynamic correction, and 3DVAR data assimilation. It systematically verified the core hypothesis: the multi-source optimization method can significantly improve the simulation accuracy of the wind field in forest areas, and the optimization effect follows clear physical laws. Regarding four research questions, the results show that (i) the YSU scheme performs best in spring unstable stratification and winter upper layers, while the MYJ scheme has a slight advantage in the near-ground layer but is not significantly different from the YSU scheme, revealing the seasonal and height patterns of the PBL scheme; (ii) local LAI correction reduces the error of near-ground layer wind speed by 35% and improves the wind direction accuracy by 28%, and the optimization effect decays with height and strengthens with increasing LAI. Local LAI correction should be regarded as a standard configuration for forest WRF simulation, especially suitable for summer simulations with high LAI values; (iii) 3DVAR assimilation achieves the highest wind speed correlation coefficient of 0.869 and wind direction of 0.813, confirming the necessity of the season-specific B matrix, and it should be constructed separately for different seasons instead of using a single B matrix; (iv) multi-source optimization significantly improves the wind energy evaluation indicators. The error of the c value at 50 m in spring is reduced by 78.6% compared to the traditional WRF model, verifying the engineering value of the multi-source optimization framework of this study for wind energy resource assessment. The core innovation of this study lies in clarifying the synergy effect of PBL optimization selection, LAI local dynamic correction, and 3DVAR assimilation, revealing the vertical and seasonal physical laws of wind field simulation in forest areas, enhancing the applicability of the WRF model in complex vegetation coverage and non-flat terrain conditions, providing scientific basis for forest wind power development, and offering a methodology that can be extended to other complex vegetation areas. Future research can further integrate seasonal variations and complex weather patterns to achieve more accurate simulation of turbulent structures.