Performance of the CMA-GD Model in Predicting Wind Speed at Wind Farms in Hubei, China

The CMA-GD model has demonstrated robust predictive capacities in forecasting various weather phenomena such as precipitation, typhoon track and intensity, as well as cold air outbreaks in South China ^[1-3]. The Guangzhou Institute of Tropical and Marine Meteorology has carried out a series of research on the application, evaluation, and improvement of the CMA-GD model, such as the evaluation and improvement of the CMA-GD model's wind field forecast in South China by using terrain parameterization technology. CMA-GD model shows substantial wind deviation on the complex terrain of South China ^[4-8]. CMA-GD with SSOP scheme reduces the wind deviation in the lower troposphere and dramatically alleviates the ground wind speed deviation ^[9-14]. By using the ground observation and the observation data of two sounding stations in South China ^[15-18], the distribution and daily change of simulated wind in the Pearl River Delta region, southern Jiangxi Province, and the coastal area of South China of CMA-GD model are tested ^[19-20]. It shows a large number of stations with opposite surface wind direction in Eastern Guangxi and Southern Jiangxi during the night to morning, and the overall overestimation of surface wind in the coastal area in the afternoon ^[21-24]. The comprehensive observation and research on the warm area rainstorm (WSTR) in South China were carried out, the ability and limitations of CMA-GD to simulate the extreme rainfall in the warm area of complex terrain were analyzed ^[25-30], and the characteristics and weather environment of extreme rainfall events in the warm area of South China were comprehensively studied ^[31-32]. Based on the hourly ground observation data from 2015 to 2019, the characteristics of rainfall from night to morning in the warm season in South China are studied ^[33].

Some studies have also examined other numerical forecast models. For example, He et al. used the WRF model to downscale three commonly used global circulation forecast fields (ECMWF, GFS, T639) at a horizontal resolution of 10km for four typical months in the Chinese region ^[34]. The forecast performance of 70m wind speeds within 24 hours with observed data from 400 wind measurement towers in China were also compared ^[34-40]. Xia et al. employed probabilistic statistical methods to evaluate the mean absolute error, standard deviation, correlation coefficient, and accuracy of wind speed forecasts from ECMWF, GRAPES (CMA-GD), and JAPAN models for 84 sites within Guizhou Province ^[41]. Shi et al. evaluated the continuous trend and error distribution of 9km × 9km wind speed forecast products from the BJ-RUC model using observed hourly wind speed data at hub height from a wind farm in Inner Mongolia, using methods such as trend analysis, error and bias frequency, and error phase inversion ^[42].

Although the above works have carried out extensive and in-depth research on CMA-GD and other models from all sides, there is still a lack of application evaluation of the CMA-GD model in wind farms. Especially in wind farms with complex mountainous terrain in Hubei Province, there is a lack of in-depth analysis and research on the performance of CMA-GD in wind farms. For example, the root mean square error, correlation, average absolute error, and other performance indicators of wind speed prediction are not clear, and the performance comparison with other prediction models (such as the EC model) is not clear.

In recent years, China's wind power generation has grown rapidly. By 2021, the installed capacity of wind power connected to the grid ranks first in the world, accounting for about 13% of the total installed capacity of China's power supply, and the power generation capacity accounts for about 7.5% of the total social electricity consumption. Wind power generation occupies an increasing proportion of the energy industry, which promotes the sustainable development of energy and the economy ^[43]. Large-scale wind power grid connection has caused a strong impact on the power grid. Accurate wind speed prediction can effectively alleviate this problem. This prediction is challenging due to the irregular nonlinear distribution of wind and chaotic dynamic characteristics ^[44-45]. Therefore, it is essential to systematically evaluate the wind speed prediction performance of CMA-GD for wind farms. On the one hand, it can provide a reference for better use of model products and further improvement of modeling technology; on the other hand, it also has significant economic and social benefits.

The purpose of this study is to evaluate the prediction performance of the CMA-GD model in the wind farm in Hubei. This study mainly includes the following aspects:

(1) What is the performance of the CMA-GD model in forecasting wind farms in Central China, especially in mountainous areas of Hubei Province?

(2) What are the observed and simulated wind distribution characteristics at the hub height of the wind turbine?

(3) Compared with the EC model, what is the performance of the CMA-GD model?

(4) Analyze the possible causes of errors and suggestions for further improvement.

The structure of the paper is as follows. Section 2 gives the data and model description, section 3 presents the performance and evaluation results of the wind speed forecast, and section 4 includes the conclusion and discussion.

In this study, the performance of the CMA-GD and EC models in predicting wind speeds in two wind farms located in Huashan (Wind farm A) and Tianhekou (Wind farm B) in Suizhou Municipality, Hubei Province was evaluated using hourly observation data from the wind measuring tower at the height of 70m (wind turbine hub height) in 2019. Abnormal observation data, including those that arose from equipment faults and icing, were excluded from the analysis. However, it should be noted that the wind observation data for Wind Farm B was not available for January, February, and March.

The topographic and geomorphological maps of Farm A and Farm B are illustrated in Fig. 1 and Fig. 2, respectively. The wind measurement tower in Wind Farm A is located at an altitude of 450m, situated in the middle of a relatively high mountain ridge, with a significant altitude difference of approximately 300m compared to the surrounding area. In addition, the wind measurement tower in Wind Farm B is positioned at an altitude of 330m, near the edge of a low mountain, with a relatively smaller altitude difference of around 100m. It is evident from these observations that the terrain complexity of Wind Farm A is substantially higher than that of Wind Farm B.

Figure 1.  The geographic location of the wind measurement tower in Wind Farm A.

Figure 2.  The geographic location of the wind measurement tower in Wind Farm B.

CMA-GD adopts the CMA-GD-GZ model, which is based on the CMA-GD non-static mesoscale model. The semi-implicit semi-Lagrangian time difference scheme is used for the grid design of longitude-latitude grid points. Araka-C grid is adopted in the horizontal direction, the Chaney-Philips vertical layering scheme is adopted in the vertical direction, and the vertical coordinate is the altitude terrain tracking coordinate. The Helmholtz equation is employed for an implicit solution. The model incorporates various schemes for physical processes, including the RRTMG long and short wave radiation scheme, WSM6 cloud microphysics scheme, NSAS convection parameterization scheme, SMS land surface parameterization scheme, and NMRF boundary layer scheme. The range of the prediction test model is 102º– 122.04ºE, 23º–38.99ºN, the horizontal grid distance of the model is 0.03º×0.03º, the vertical direction is 65 layers, and the time integration step is 60s. The initial field and side boundary conditions adopted by the model are the EC analysis field and the prediction field every 6 hours. The horizontal resolution of EC data is 0.1º×0.1º. The integration time of the mode is 12:00 on January 1, 2019 (universal time). The simulation time limit is 72 hours per day, and it is output every 1 hour. The integration time is 1 year.

In order to conduct a wide range of performance comparisons, three indicators, namely root mean square prediction error (RMSE), mean absolute error (MAE), and correlation coefficient, are used to evaluate the performance of CMA-GD and EC prediction models ^[18], which are described as follows:

The RMSE formula is as follows:

MAE formula is as follows:

The correlation coefficient is calculated as follows:

where O_i represents the observed value, F_i represents the predicted value, n represents the total number of samples, O_i represents the average value of the observed samples, and F_i represents the average value of the predicted samples.

In this study, Wind Farm A and Wind Farm B are selected to compare and analyze the prediction results of CMA-GD and EC models. Excluding the months when the data quality of the wind tower equipment is not high due to icing and equipment failure, Wind Farm A and Wind Farm B have 12 months and 9 months of wind measurement and observation data with a height of 70 m, respectively. The prediction and comparison results in Wind Farm A are shown in Table 1 and Fig. 3.

Figure 3.  Comparison of CMA-GD and EC root mean square error.

In Wind Farm A, the CMA-GD model has a better prediction effect than the EC model for the whole year. The annual RMSE index is 0.11 m s^–1 smaller than EC, the MAE index is 0.10 m s^–1 smaller than EC, and the correlation coefficient R is 0.05 higher than EC. However, compared with the EC model, the error in February and March is relatively large. It is found that the CMA-GD model may take more consideration of tropical climate conditions and lack of consideration of the climate conditions in the middle and lower reaches of the Yangtze River. Hence, the prediction performance of the model is slightly poor in the cold season in Hubei.

In order to deeply analyze the wind forecast effect, the data of typical months (January, April, July, and November) are selected to draw the rose chart of the CMA-GD forecast model, EC forecast model, and the actual observed wind direction of the Wind farm, as shown in Fig. 4 to Fig. 7.

Figure 4.  Wind rose in January.

Figure 5.  Wind rose in April.

Figure 6.  Wind rose in July.

Figure 7.  Wind rose in October.

It can be seen from Fig. 4 that the wind farm has a typical winter climate in the Hubei mountainous area in January, and the dominant wind direction of the Wind farm is NE. CMA-GD and EC can accurately predict the dominant wind direction. It is found that the monthly average wind direction of CMA-GD is 88.19 degrees, the monthly average wind direction of EC is 80.20 degrees, and the actual observed wind direction is 94.34 degrees. The wind direction forecast of CMA-GD is closer to the actual observation value, and the wind direction forecast error of EC is slightly larger.

Figure 5 shows the wind direction prediction ability of the two prediction models in April. CMA-GD and EC show relatively close wind direction prediction ability. The dominant wind direction in the whole month is still NE. In terms of the prediction frequency of the NE wind direction, the two prediction models are less, and the wind direction prediction is slightly north. The secondary dominant wind direction in this month is SE, but the predictions of the two prediction models are close to the SW direction.

It can be seen from Fig. 6 that in July, the wind direction prediction errors of CMA-GD and EC prediction models are too large. The observed monthly average wind direction is 197.86 degrees, the monthly average wind direction of CMA-GD is 122.44 degrees, and the monthly average wind direction of EC is 125.31 degrees. In this month, EC has slightly better wind direction prediction ability than CMA-GD.

Figure 7 shows that there is a large difference between the distribution of the predicted and measured wind directions of CMA-GD and EC in October. It can be seen from Fig. 7 that the prediction of wind direction also directly affects the accuracy of wind speed prediction. In October, the root mean square error of CMA-GD and EC prediction models reached 3.40 m s^–1 and 3.54 m s^–1, respectively.

In order to compare the performance of CMA-GD and EC prediction models more widely and fairly, we chose Wind Farm B for comparative analysis. Due to the influence of equipment failure and network communication, the following periods with missing data and poor quality are excluded, as shown in Table 2.

In Wind Farm B, the comparison results of root mean square prediction error (RMSE), mean absolute error (MAE), and correlation coefficient of CMA-GD and EC prediction models are shown in Table 3.

On the whole, the prediction performance of the CMA-GD model is worse than that of the EC model. The root mean square error is 0.47 m s^–1 higher, the average absolute error is 0.46 m s^–1 higher, and the correlation coefficient is 0.08 lower. The prediction error of the CMAGD model from June to August is close to that of the EC model, and the prediction error of other months is larger than that of the EC model. From the forecast trend, the CMA-GD wind speed forecast is consistent with the measured wind speed trend, but there is a large systematic deviation in the forecast results. For example, the maximum wind speed predicted by the CMA-GD model in July is only about 8 m s^–1, which is obviously small, while the EC model forecast is closer to the measured wind speed forecast, and the most typical wind process can be reported.

The following four months, April, July, September, and December, are selected for a more detailed comparative analysis of wind speed forecast. As shown in Fig. 8–11.

Figure 8.  Comparison of wind speed forecast of Wind Farm B in April 2019.

Figure 9.  Comparison of wind speed forecast of Wind Farm B in July 2019.

Figure 10.  Comparison of wind speed forecast of Wind Farm B in September 2019.

Figure 11.  Comparison of wind speed forecast of Wind Farm B in December 2019. Data missing from 14:00 December 23 to 16:00 December 24.

It can be seen from Fig. 8 that there are many gale days in April. The EC prediction model can accurately predict the changing trend of wind speed, and the phase and fluctuation amplitude of wind speed can be accurately captured, showing the prediction ability consistent with the actual observation. Although the CMA-GD forecast model is relatively consistent in the wind speed change trend, the forecast ability on typical gale days is significantly insufficient, and there are systematic deviations. For example, the gales from April 16 to April 21 and from April 26 to April 29 can not be accurately predicted. Most of the CMA-GD forecasts are below 6 m s^–1, and most of the measured wind speeds are above 8 m s^–1.

Figure 9 shows that the wind speed in July is lower than 6 m s^–1, and the trend of CMA-GD and EC prediction models is consistent, but the CMA-GD in the high wind speed section greater than 6m s^–1 is obviously smaller. The correlation coefficients r of CMA-GD and EC are close, 0.56 and 0.57, respectively. The monthly root mean square error of CMA-GD is 2.45 m s^–1, that of EC is 2.14 m s^–1, and that of CMA-GD is 0.31 m s^–1 higher than that of EC.

Figure 10 shows that the two prediction models of large wind speed forecast from September 14 to 16 are smaller, but the EC forecast is closer to the measured wind speed than the CMA-GD forecast. The monthly root mean square error of CMA-GD is 2.33 m s^–1, that of EC is 1.86 m s^–1, that of CMA-GD is 0.47 m s^–1 higher than that of EC, that of CMA-GD is 1.92 m s^–1, that of EC is 1.46 m s^–1, and that of CMA-GD is 0.46 m s^–1 higher than that of EC.

Figure 11 shows the comparison of wind speed forecasts of Wind Farm B in December. It can be seen from the comparison chart that from December 8 to 10, the predicted wind speed of EC was significantly larger, and the predicted wind speed of CMA-GD was closer to the measured wind speed. From December 14 to 19, the predicted wind speed of CMA-GD was significantly smaller. The root mean square error of CMA-GD is 2.58 m s^–1, and that of EC is 1.93 m s^–1, which is higher than that of EC by 0.65 m s^–1. The average absolute error CMA-GD is 2.24 m s^–1, EC is 1.62 m s^–1, and CMA-GD is 0.62 m s^–1 higher than EC.

The following conclusions and suggestions are obtained by comparing the measured wind speeds of the two wind farms in Hubei in 2019 with the CMA-GD and EC prediction models:

1) In both real-world conditions and numerical model simulations, terrain plays a significant dynamic role in near-surface wind fields. The more rugged and complex the terrain, the higher the demand for resolution in the model′s terrain data. In the more rugged terrain of Wind Farm A, the wind speed prediction error of the CMA-GD model is smaller than that of the EC model. In the flatter terrain of Wind Farm B, the wind speed prediction error of the CMA-GD model is larger than that of the EC model. The reasons for these differences may be due to deviations between the actual terrain and the modeled terrain. The width of the ridges where Wind Farm A and Wind Farm B wind measurement towers are located is approximately 300m and 200m, respectively, both of which are smaller than the spatial resolution of the current operational numerical weather prediction models. Therefore, the models can only approximate the real terrain by using higher-resolution terrain data. The spatial resolution of the CMA-GD model is 0.03º×0.03º, so in the more complex terrain of Wind Farm A, the deviation between the model terrain and the actual terrain is smaller than that of the EC model with a spatial resolution of 0.1º×0.1º. This results in a smaller prediction error of the wind speed for the CMA-GD model compared to the EC model. Therefore, in complex mountainous areas, it is recommended to use the CMA-GD model with higher spatial resolution. In the future, it is suggested to further reduce the spatial resolution of the CMA-GD model to reflect the complex terrain of mountainous areas better and improve the accuracy of wind speed forecasting.

2) In the prediction of the two Wind farms, the prediction performance of CMA-GD in the cold season is lower than that of the EC prediction model. In summer, the prediction performance of CMA-GD is better than or close to that of the EC prediction model, which may be related to the integration of more tropical meteorological observation data with this prediction model. It is suggested that the CMA-GD model assimilates more observation data of Hubei region, such as wind measurement observation data at different heights of the Wind farm and high-altitude wind measurement observation data.

3) At present, the time resolution of the output results of the CMA-GD prediction model is 1 hour, and the prediction time limit is 72 hours in the future. There is still a gap between the time resolution of 15 minutes required by the Wind farm and the prediction time limit of 240 hours in the future. It is suggested to improve the time resolution and prediction time limit of the CMA-GD prediction model.