Fengyun-4 Geostationary Satellite-Based Solar Energy Nowcasting System and Its Application in North China

To achieve carbon peak and neutrality goals, China will increase the share of renewable energy in its energy structure to reduce carbon emissions (Shi et al., 2021; Li et al., 2022). China has the world's largest installed photovoltaic (PV) capacity and growth rate (IRENA, 2020). Until August 2021, China had an installed capacity of 275.1 GW and had surpassed its 2016 capacity by 255.4%. The ratio of installed capacity to China’s full-caliber installed capacity has increased from 4.7% (2016) to 12.0% (August 2021)^①. The share of PV energy in China's energy structure will undoubtedly increase significantly in the coming decades (Burandt et al., 2019; Prăvălie et al., 2019; Li et al., 2020; Bai et al., 2021). The intermittence and instability of solar energy, however, pose a major challenge for the integration of solar PV into the energy grid. Solar energy mainly depends on the availability of surface solar irradiance (SSI), which is highly susceptible to variations in atmospheric components (clouds, aerosols, water vapor, etc.). This results in significant fluctuations in the PV output power. Therefore, accurate forecasting of SSI fluctuations is of great importance for utilizing and developing solar PV and reducing carbon emissions (Lamsal et al., 2018; Senatla and Bansal, 2018).

The forecast method for SSI is generally divided into three categories. Years- and decades-ahead predictions are mainly made using climate models that are commonly used for solar PV plant deployment and planning (Wild et al., 2015; Zou et al., 2019). For short-term (day/hour-scale) forecasts, regional numerical weather prediction (NWP) simulations are advantageous; the results obtained through these are useful for scheduling and optimizing microgrids and grid-connected power allocation for the next 24–48 hours (Antonanzas et al., 2016; Razagui et al., 2021). For nowcasting (0–3 h ahead), satellite-based methods are widely used, which is important for the safe operation, control, and management of PV plants and is also a prerequisite for grid connection of PV plants (Zhu et al., 2019).

The satellite-based SSI nowcasting method consists of two components. The first is to estimate SSI as accurately as possible using satellite reflectance measurements. The second is to forecast SSI at the next target time, which requires estimation of the cloud motion vector (CMV). The diverse methods used to estimate SSI from satellites can be divided broadly into physical and semi-empirical methods; both have been widely used for different satellites (e.g., NOAA series, Meteosat Second Generation (MSG), and Himawari-8) (Kallio-Myers et al., 2020; Peng et al., 2020; Letu et al., 2021). The physical method uses atmospheric and surface properties retrieved from satellite measurements and other sources, such as reanalysis products, to drive a radiative transfer model to calculate SSI. This method is based on solid physical foundations, and its uncertainty depends heavily on the availability and accuracy of the input variables. Additionally, this method requires more computing power than empirical methods. The semi-empirical method estimates SSI directly from satellite observations using a simple but effective method. One representative method of this kind in the solar energy community is the semi-empirical method for MSG, called the Heliosat method. This method was originally developed in the last century and has greatly facilitated research and application of solar energy in Europe (Cros et al., 2004; Rigollier et al., 2004). The basic principle is very simple, but its operational implementation is delicate and requires a fair amount of site-specific accounting (Kleissl, 2013). In recent years, machine learning, which can be classified as an empirical method, has developed rapidly for SSI estimation (Damiani et al., 2018; Peng et al., 2020). Empirical methods mainly use satellite observation and generally do not need atmospheric and surface properties. Therefore, they can be extremely useful when atmospheric and surface properties are not reliably available or have large uncertainties (Huang et al., 2019). Another advantage of this method is that SSI estimations can be obtained in the native spatial resolutions of the original satellite measurements. This is very useful for solar energy forecasting because the higher the spatial resolution, the better is the application.

With regard to CMV derivation, three methods are widely used: phase correlation, optical flow, and block-matching (Amillo et al., 2014; Kallio-Myers et al., 2020; Yang et al., 2020). The phase correlation method obtains motion information from the phase difference of consecutive fields. The basic principle is the shift property of the Fourier transform, which states that translation of pixels in the spatial domain produces proportional phase shifts in the frequency domain (Wang et al., 2018). The optical flow method derives the CMV based on the image gradient in which three assumptions are made. First, the pixel brightness of the image is constant between the two frames. Second, the motion in the image varies little over time. Third, the same or similar motion is shared by the target pixel and its neighboring pixels, i.e., the spatial smoothing is constrained (Nonnenmacher and Coimbra, 2014). The block-matching method assumes that cloud structures are constant along motion trajectories over a short time period. The CMV is calculated through the cloud structure displacement obtained from consecutive images (Hammer et al., 1999). CMV is generally obtained by directly using 2–3 images from geostationary satellites taken within a short period of time. Retrievals of cloud physical properties, such as cloud optical thickness, cloud top height, and cloud effective radius, have been shown to improve the CMV derivation (Wang et al., 2019).

SSI nowcasting can be achieved using a combination of CMV information and SSI estimates. Numerous SSI nowcasting studies have been conducted based on satellite sensors (e.g., Spinning Enhanced Visible and Infrared Imager (SEVIRI)) (Arbizu-Barrena et al., 2017; Gallucci et al., 2018; Mouhamet et al., 2018; Wang et al., 2019). Fengyun-4A (FY-4A), launched by China in 2016, is a new generation geostationary meteorological satellite. The advanced geosynchronous radiation imager (AGRI) on board the FY-4A can provide images with much higher spatial, temporal, and spectral resolutions than previous imagers. It has great potential for SSI nowcasting, which still needs to be explored (Zhang et al., 2020; Xian et al., 2021). There have been a few attempts, but all were limited in their spatial and temporal coverage and satellite-based detection of cloud motion (Yang et al., 2019; Chen et al., 2020; Yang et al., 2020). More importantly, many issues still require further investigation.

In this study, we explore the potential of FY-4A in SSI nowcasting. We attempt to answer the following three scientific questions: 1) Can the clear-sky model be used to accurately estimate clear-sky SSI even under heavy pollution conditions. 2) what is the accuracy of SSI estimation expected by using an empirical method to the AGRI onboard the FY-4A, and how can its accuracy be improved. and 3) Can the empirical method be extended to DNI estimation, and what is the expected accuracy. China has a vast absolute area of DNI superb potential [>200 000 km², (Prăvălie et al., 2019)], which means that accurate estimates and nowcasting of DNI fluctuations are extremely important for the development of solar PV energy in China. Therefore, it is of great significance to explore FY-based DNI estimates and the nowcasting method.

This paper is organized as follows: Section 2 describes the clear-sky model, SSI estimates model, nowcasting algorithm, and data used in the study. Section 3 presents the results of the evaluation of FY-4A SSI estimates and forecast products. The discussion and conclusions are presented in sections 4 and 5.

FY-4A is located over the equator at a longitude of 104.7°E. The AGRI onboard the FY-4A takes satellite images of China and its surrounding areas with varying spatial (from 0.5 km to 16 km) and temporal (from 15 min to 2 h) resolutions. To simplify the method and thereby fulfill the operational usage of the method, we only used Level 1 spectral reflectance at the near-infrared channel (0.83 µm) with a spatial resolution of 1 km to calculate the cloud index (CI) and CMV information. The data are available online (http://satellite.nsmc.org.cn/portalsite/default.aspx).

The Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) uses the Goddard Earth Observing System Version 5 data assimilation system. MERRA-2 provides long-term (from 1980 to the present) radiation, meteorological, and atmospheric aerosol products, which have been widely used to study atmospheric environments, climate change, and solar energy (Randles et al., 2017). The hourly MERRA-2 aerosol optical depth at 550 nm (AOD₅₅₀), Ångström exponent (AE), precipitable water vapor, ozone amount, and surface albedo products (http://disc.sci.gsfc.nasa.gov) are used to drive a parametrization radiative transfer model to calculate clear-sky SSI.

The SSI measurements at Xianghe (39.75°N, 116.95°E, Fig. 1) in 2018 were used for the evaluation. GHI and DNI were measured using a Kipp and Zonen CM21 pyranometer and CHPI pyrheliometer, respectively (Liu et al., 2021). The SSI measurements were quality controlled through procedures recommended by the Baseline Surface Radiation Network, with uncertainties of approximately 6% and 3% for GHI and DNI, respectively. Only quality-assured measurements were used for the evaluation, which led to different sample numbers of GHI and DNI. Xianghe is located in the Beijing–Tianjin–Hebei region, which is a typical polluted region in China. This region has been gradually increasing the share of solar PV energy in its total energy consumption to balance energy requirements with environmental protection.

1-minute SSI measurements were used to detect cloud occurrence using a random forest method (Liu et al., 2021). Daily cloud occurrences were used to classify each day into four distinct categories: C1: 0%–10%; C2: 10%–50%; C3: 50%–90%; C4: 90%–100%, roughly corresponding to clear, moderately clear, cloudy, and overcast conditions, respectively. The SSI estimates and nowcasting were evaluated under these four sky conditions to investigate their potential different performances.

Figure 2 presents the monthly occurrence frequency of these four sky types in 2018. The cloud occurrence frequency showed a seasonal pattern. Skies were much clearer in winter than in other seasons, while overcast skies dominated in summer, especially in July when 27 days were overcast.

Figure 1.  (a) Geographical location of Xianghe observation site, and (b) the instrument.

Figure 2.  Monthly proportion of four sky conditions at Xianghe in 2018.

Clear-sky SSI estimation is a critical step in empirical SSI algorithms. In this work, clear-sky SSI was calculated using the REST2v5 model because of its excellent performance for supporting SSI estimates and nowcasting in a worldwide test under various conditions (Antonanzas-Torres et al., 2019; Sun et al., 2021). REST2 splits the shortwave spectrum into two bands (290–700 nm and 700–4000 nm), and DNI and diffuse horizontal irradiance (DHI) are parameterized for each band (Gueymard, 2008). The GHI is calculated as the summation of the DNI multiplied by the cosine of the solar zenith angle and DHI.

The all-sky SSI values were obtained by multiplying the REST2 clear-sky SSI calculation and the clear sky index ($ {K_{\text{c}}} $: the ratio of observed all-sky SSI to the expected clear-sky SSI). $ {K_{\text{c}}} $ is parameterized as a function of the CI:

where $ {n_t}(i,j) $ denotes the CI derived from FY-4A reflectance measurements. The relationship between CI and $ {K_{\text{c}}} $ is determined based on the historical data of Xianghe station.

The CI at a given pixel $ (i,j) $ and time $ t $ is calculated using Eq. (2):

where $ {\rho _t}(i,j) $ is the FY-4A observed reflectance, and $ {\rho _{t{\text{,c}}}}(i,j) $ and $ {\rho _{t,{\text{g}}}}(i,j) $ represent the reflectance associated with the brightest clouds and clear skies within a given period, respectively. The operational advantage of this all-sky SSI estimation method is that it does not require precise satellite calibration information. As the method is self-calibrating, the upper and lower bounds of the dynamic range at the target pixel are established from the data history (Kleissl, 2013). The upper bound of the dynamic range, $ {\rho _{t{\text{,c}}}}(i,j) $, represents the overcast condition with the brightest clouds. We defined $ {\rho _{t{\text{,c}}}}(i,j) $ as the 95th percentile of all reflectance values throughout the year in the target region (latitude 36°N to 44°N and longitude 112°E to 120°E), which considers potential measurement uncertainty. Given $ {\rho _{t,{\text{g}}}}(i,j) $, the lower bound of the dynamic range is highly dependent on surface reflectance, which shows remarkable temporal variation; it is estimated by an iteration procedure for each month. First, all satellite measurements of reflectance for each pixel in a given month are averaged as an initial threshold value ($ {\rho _{{\text{thres}}}} $). The new average is then calculated from the reflectance smaller than the sum of the initial threshold and a specified bias $ \varepsilon $, which is defined as 0.035 $ {\rho _{t{\text{,c}}}}(i,j) $ here. This new average is used for the next iteration. Then, the iteration proceeds until no reflectance exceeds the threshold by a magnitude of $ \varepsilon $, and the last-iterated $ {\rho _{t,{\text{g}}}}(i,j) $ is used as the lower bound.

Note that the FY-4A-measured clear-sky reflectance shows a convex parabolic variation with the solar zenith angle (SZA) over the daytime. This means that $ {\rho _t}(i,j) $ at local solar noon would exceed $ {\rho _{t,{\text{g}}}}(i,j) $. Then, the $ {n_t}(i,j) $ obtained based on Eq. (2) would tend to identify the clear-sky pixels at local solar noon as cloud pixels, eventually leading to much lower calculated clear-sky SSI values than actual SSI values. Therefore, this issue was addressed by introducing a parameterization of the FY-4A clear-sky reflectance to SZA, as follows:

where $ {\rho _{\text{c}}} $ denotes corrected reflectance, $ {\rho _0} $ denotes FY-4A clear-sky reflectance, $ \theta $ represents SZA.

The corrected reflectance and corresponding dynamic range are given in Fig. 3b.

Figure 3.  (a) The function relation between FY-4A reflectance and the cosine of SZA under the clear-sky condition, and (b) the dynamic range for Xianghe station from the near infrared channel of the FY-4A satellite in 2018.

The GHI is given by Eq. (4):

where $ {\text{GH}}{{\text{I}}_{{\text{cs}}}} $ represents the GHI under clear sky calculated by the REST2 model.

The DNI is given by Eq. (5) (Pfeifroth et al., 2019), which is determined based on the historical data of Xianghe station:

where $ {\text{DN}}{{\text{I}}_{{\text{cs}}}} $ represents the DNI under clear sky, which is also calculated using the REST-2 model.

The forecast method includes two operational steps: forecasting clear-sky SSI background at the target horizons, and superimposing cloud attenuation ($ {K_{\text{c}}} $) on the background. We assumed that all parameters, except the SZA, are held constant with a forecast horizon up to 6 h. The present values (0 min ahead) provided by MERRA-2 are used as the input predicted values for the prediction of the clear-sky SSI background. The SZA is the main cause of the alteration of clear-sky SSI background. The SZA values at target horizons are obtained according to the time and geographical location information. The FY-4A satellite images were first converted to CI images using the abovementioned empirical method. The CMV was subsequently calculated using the block-matching method and consecutive CI images with time intervals of 20 min. Based on the CMV information, the CI and $ {K_{\text{c}}} $ of the forecast horizon were obtained, and the SSI nowcasting was finally completed.

The development of clouds is not determined by dissipation and formation for the forecast time of several hours, and cloud tracking methods (e.g., the block-matching method) using images from geostationary satellite show good performance. As a demonstration, we used a rectangular area [112°–120°E, 36°–44°N] as our research subject. The images were divided into multiple square boxes (0.5° × 0.5°), and the motion vectors for each square region were derived using the following procedures. The block-matching method identifies cloud structures according to the pixel intensities in consecutive CI images. Assuming that pixel intensities are constant along motion trajectories over a short time period, the CMV of each square area was subsequently calculated using the cloud structure displacement (Hammer et al., 1999).

According to the assumption of constant pixel intensities, the vector v used to describe the motion is given by:

where n_i denotes the CI at position x of the image at time t_i. The actual and past images were subscribed with 1 and 0, respectively.

The best vector describing the cloud motion was determined using the displacement vector with the minimum mean square pixel difference (MSE):

The accuracy of the SSI estimates and nowcasting with 20-min temporal resolution was evaluated for the Xianghe observations. For spatial matching, we used the SSI estimates and nowcasting for the closest pixel to the Xianghe station. The FY-4A SSI estimates and nowcasting performance were evaluated using the following statistical indices: root-mean-square error (RMSE), mean bias error (MBE), and their normalized counterparts (nRMSE and nMBE).

In this section, we first evaluate the performance of the REST2 model, then the accuracy of the estimated GHI and DNI, and finally, the forecast accuracies for time horizons between 30–180 min. All evaluations are detailed under the four sky conditions.

Figure 4 shows the scatter statistics of the clear-sky GHI and DNI estimates based on the REST2 model. The GHI estimates were very close to the ground-based measurements, with nRMSE and nMBE of 4.93% and 1.68%, respectively. The REST2 model tended to slightly overestimate GHI on clear-sky days, and it tended to overestimate DNI by a larger magnitude compared to GHI, with an nMBE value of 8.67%. Large overestimations of DNI were observed in low and moderate DNI conditions. One explanation is that a lower DNI is associated with a high aerosol loading, which is generally underestimated by the MERRA-2 reanalysis (Huang et al., 2021). Notably, the model estimates were very close to the surface measurements when DNI was high. The high output power of solar PV plants depends on high DNI values. Compared to the low and moderate DNI values, the fluctuations of high DNI values will induce heavier impact on the safe operation, control, and management of solar PV plants. Therefore, the accurate calculation of high DNI values is crucial for the utilization of solar PV and is the focus of DNI calculations. Overall, the REST2 model performed well in supporting the GHI and DNI estimates and nowcasting evaluations at the Xianghe station under different conditions.

Figure 4.  Density plot of REST2 model GHI (a) and DNI (b) versus surface measurements under clear-sky conditions. The black line denotes a 1:1 line. The red lines indicates linear regression results.

Figure 5 presents the density plot of FY-4A GHI and DNI under all-sky conditions against surface observations at Xianghe. The estimated GHIs show good correlation with the surface measurements, with nRMSE and nMBE of 22.4% and 5.9%, respectively. The percentage of the estimated GHI values exceeding the measurements by 100 W m^–2 was approximately 15.9%. These overestimated GHI values were concentrated at 200–500 W m^–2, comprising the dominant contribution to GHI overestimation. Accurate calculation of high GHI values is crucial for solar energy estimation and forecasting. The good performance of the FY-4A GHI estimation model when the GHI is high indicates that the model can support FY-4A GHI estimation and nowcasting under all-sky conditions. Compared with the estimated GHI, the performance of the estimated DNI is worse, and the model needs to be improved in this aspect. In general, the FY-4A DNI was overestimated with an nMBE value of 24.1% (Fig. 5b). The empirical-cloudy model performs well with a high DNI condition and tends to largely overestimate DNI when DNI values are lower than 700 W m^–2. One explanation is that if the ground station is covered by the shadows of clouds or surrounding features (namely, the 3D effect), its real-time measured DNI will be lower than that of other locations within the matching pixel in satellite images, thus the measured DNI values will be lower than satellite-derived values (Huang et al., 2019; Jiang et al., 2019). The deviation of clear-sky DNI background and CI is also a contributor to the overestimation.

Figure 5.  Density plot of the SSI estimation model GHI (a) and DNI (b) versus SSI measurements under all-sky conditions in 2018. The black line denotes a 1:1 line. The red lines indicate liner regression results.

Table 1 shows the statistical indices of the estimated GHI and DNI in this study compared with those in previous studies based on empirical methods. The nRMSE and nMBE of the GHI estimates in Shanghai obtained by Chen et al. (2020) were 34.7% and 11.4%, respectively, which were higher than those obtained in our work by 12.3% and 5.5%. The nRMSE values of the estimated GHI and DNI in our study were lower by 3.5% and 3.9%, respectively, compared to those obtained by Jia et al. (2021) in their study of northern China. The improvement of our estimation model eliminates the bias related to SZA variation by introducing a parameterization of the FY-4A clear-sky reflectance to SZA.

The evaluation of the estimated GHI and DNI for different sky conditions is shown in Fig. 6. The accuracies of GHI and DNI estimations declined with an increase in cloud occurrence (from clear-sky C1 to overcast C4). GHI estimates feature less deviation under clear, moderately clear, and cloudy conditions. The MBE values of GHI varied from –22.4 W m^–2 to 23.5 W m^–2, and the RMSE values increased from 41.1 W m^–2 to 90.5 W m^–2 as the sky conditions varied from C1 to C3. The estimation accuracy declined sharply under C4 conditions, with the MBE and RMSE reaching 62.6 W m^–2 and 144.5 W m^–2, respectively. The performance of the estimated DNI was significantly worse than that of the GHI estimation, featuring larger deviations under all four conditions. The MBE values varied from –9.1 W m^–2 to 173.1 W m^–2, while the RMSE values increased from 94.2 W m^–2 to 253.7 W m^–2. The poor performance under the C4 condition had a limited impact on the PV plants because the grid-connected power allocation can be adjusted to suit C4 conditions based on regional numerical models. Therefore, the performance of the GHI and DNI estimates under other sky conditions, especially C1 and C2, is the focus of GHI and DNI estimation and nowcasting and is crucial in the utilization of solar PV.

Figure 6.  Density plot of the SSI estimation model GHI (a–d) and DNI (e–h) versus observed SSI under different sky conditions. The black line denotes a 1:1 line. The red lines indicate linear regression results.

To characterize seasonal variations in model accuracy, the monthly mean estimated GHI and DNI and surface measurements were compared in Fig. 7. The monthly mean estimations of GHI and DNI were larger than the surface measurements. The difference between model and observation presented a seasonal pattern with a large discrepancy in spring–summer and a small difference in autumn–winter. The seasonal pattern of the GHI and DNI model discrepancy is related to the seasonal variation in cloud occurrence. In spring and summer, the proportion of days with the C4 condition was higher than 40%, especially in July, with the proportion reaching 90% (as shown in Fig. 2). High cloud occurrence reduces the accuracy of the estimated GHI and DNI. With the decrease in cloud occurrence, the discrepancy between estimated GHI and DNI and their respective measurements was reduced in autumn and winter.

Figure 7.  Monthly mean of observed (grey line) and model (orange line) surface solar irradiance GHI (a) and DNI (b) in 2018. Shaded regions indicate one standard error of the mean.

A case of forecasted and observed CI, GHI, and DNI is presented in Fig. 8, with a 60-min forecast horizon at 0140 UTC 1 December 2018. The CI at 0240 UTC was forecasted from the CI observations and CMV information at 0140 UTC. The estimated GHI and DNI and the CI forecast were used to obtain the GHI and DNI nowcasts at 0240 UTC. The FY-4A-based GHI and DNI nowcasts and the observations had similar spatial distribution patterns. An apparent divergence between forecasts and observations occurred in the red rectangular area with broken clouds (Figs. 8a and 8d) because the CMV derivation using the block-matching method does not consider the formation and dissipation of clouds. Therefore, the CMV forecast method has relatively low predictability under the broken-cloud situation.

Figure 8.  An example of FY-4A-based (a–c) nowcasting and (d–f) observation for CI, GHI, and DNI at 0240 UTC on 1 December 2018.

The quality of GHI and DNI nowcasting within 0–3 h ahead, under all-sky conditions, is summarized in Table 2. The statistical indices for 0 min ahead were taken as deviations between the GHI and DNI estimations and the surface measurements. For 0–3-h forecast horizons, the nRMSE values of the GHI and DNI nowcasts increased slowly to 30.8% and 53.4%, respectively, surpassing those at 0 min ahead only by 8.4% and 8.0%, respectively. The nMBE values of GHI and DNI decreased with increasing forecast horizon from 5.9% and 24.1% at 0 min ahead to 0.8% and 18.6% at 3 h ahead, respectively. The GHI and DNI nowcasting methods offset the overestimation of the hybrid estimation method.

In the FY-4A satellite-based study conducted for Chengde by Yang et al. (2020), the nRMSE values of the GHI and DNI forecasts at 3 h ahead were 36.8% and 58.8%, respectively. The nRMSE of the FY-4A GHI forecasts for Shanghai calculated by Chen et al. (2020) was nearly 60% with a forecast horizon of 3 h. Both of these studies presented higher nRMSE values than the present study, suggesting that the GHI and DNI nowcasting system in this study performed better for forecast horizons up to 3 h. The CMV derivation algorithm used in the reference and our work has a similar performance (Cros et al., 2014). In this study, the correction (the parameterization of FY-4A clear-sky reflectance to SZA before CI calculation) decreases the deviations of the estimations and further improves the performance of the GHI and DNI nowcasting systems.

Figure 9 shows the nRMSE and nMBE values of GHI and DNI nowcasting under four sky conditions with forecast horizons of 0–3 h. The nRMSE values of GHI nowcasting are between 8% and 22% at 30–180 min ahead, and the increase with growing forecast horizon is relatively small under C1, C2, and C3 conditions. The nRMSE values under C4 increased rapidly as the forecast horizon increased—from 44.3% at 30 min ahead to 52.6% at 180 min. The nRMSE of DNI nowcasting exhibits similar trends, but the values are larger. The nRMSE is about 12%–13% under C1 conditions, with the 150-min forecast horizon having the smallest value. For C2 and C3 conditions, the nRMSE values increased with increasing forecast horizons—from 25.4% and 53.1% at the 30-min horizon to 33.1% and 57.9% at the 180-min horizon, respectively. Under the C4 condition, the nRMSE values of DNI show big changes over the forecast horizons of 30–180 min, ranging from 189.5% to 201.5%.

Figure 9.  The nRMSE (%) and nMBE (%) of SSI under different sky conditions with forecast horizons from 30–180 min.

The nowcasting system tends to overestimate DNI. The most severe overestimations were under C4, followed by under the C3, C2, and C1 conditions. The nMBE values of DNI showed small changes under all sky conditions, except under C4. The highest nMBE value of DNI was 31.6% at 30 min ahead under cloudy conditions. Note that the nRMSE and nMBE values of DNI nowcasting are significantly higher under C4 conditions than under other sky conditions due to the low DNI.

The benchmarks of nRMSE for GHI and DNI forecasts given by the University of Geneva are 15%–70% and 30%–160%, respectively (Beyer et al., 2009). The nRMSE values of GHI and DNI nowcasts in our study were within these ranges under all-sky conditions, except for C4. The SSI nowcasting system established in this work performed well for forecast horizons up to 180 min. The GHI nowcasting outperformed the DNI nowcasting for all forecast horizons. Furthermore, the new generation geostationary satellite FY-4A is promising for application in SSI nowcasting, but its accuracy still needs to be improved. Improving the SSI nowcasting system consists of two steps: revising the hybrid SSI estimation method and upgrading the CMV derivation algorithms.

The REST2 model performs well in estimating the clear-sky GHI in heavily polluted areas (e.g., North China). The AOD input quality in the REST2 model was key to clear-sky DNI estimation because DNI is especially sensitive to AOD compared to GHI (Gueymard and George, 2005). The empirical cloudy-sky model was first developed for MSG and is applicable to European regions (Yang et al., 2019), but it requires further revisions for it to be applicable in China using FY-4A data. The empirical coefficients of the cloudy-sky model have been calibrated based on the historical data of Xianghe station and lead to satisfactory results under clear and moderately clear conditions. The further revisions should focus on promoting the performance of the DNI estimation method under cloudy and overcast conditions. More attention should be paid to the calculation of the CI and the empirical coefficient of Eqs. (1) and (5). Long-term clear-sky DNI estimation and surface measurements should be used to fit the most suitable coefficients under different sky conditions. The empirical coefficients can be refined if high-quality SSI measurements are available across the country. Furthermore, the applicability and revision of these empirical coefficients in different regions and seasons in China is also a problem worthy of in-depth study.

The accuracy of the SSI forecast method depends on the CMV derivation algorithm. In this work, we calculated the CMV using FY-4A single-channel data. Subsequent research can develop multichannel-based CMV derivation algorithms to more accurately characterize the cloud motion. The cloud physical properties, the formation and dissipation of broken clouds, and the alteration of direction and speed in cloud motion should also be considered in future CMV derivation algorithms. For example, the square boxes used to obtain CMV fields are further divided into multiple square sub-boxes, and then intensity vectors for each sub-box, which are used to reflect the formation and dissipation process of clouds, are calculated using the alteration of pixel intensities in consecutive CI images.

A preliminary but effective SSI nowcasting system was developed based on FY-4A Level 1 reflectance data at 0.83 µm. This system was composed of an SSI estimation and forecasting method. A semiphysical SSI estimation method was developed, in which the REST2 model was first used to calculate clear-sky SSI. Subsequently, all-sky SSI was estimated according to the clear-sky SSI and the CI derived from FY-4A reflectance measurements. SSI nowcasting was developed using the CMV obtained using the block-matching method. The main conclusions are as follows:

The REST2 model performed well in supporting GHI and DNI estimates and nowcasts at Xianghe under different sky conditions. The FY-4A GHI estimates show good correlation with the surface measurements, with nRMSE and nMBE of 22.4% and 5.9%, respectively. The calibrated empirical method can be extended to DNI estimation under clear and moderately clear conditions, but it must be further improved under cloudy and overcast conditions. The FY-4A DNI estimation method tends to severely overestimate DNI when the actual DNI values are lower than 700 W m^–2. The difference between the model results and observations presents a seasonal pattern, with a large discrepancy in spring–summer and a small difference in autumn–winter. The seasonal pattern of the GHI and DNI model discrepancy is related to seasonal variations in cloud occurrences.

The nRMSE values of GHI and DNI nowcasts under all-sky conditions varied within 25.1%–30.8% and 48.1%–55.3%, respectively. The SSI nowcasting system performs well for forecast horizons up to 180 min, but its accuracy still needs improvement. The FY-4A has great potential for supporting SSI nowcasting, which would certainly enable the promotion of PV energy development and carbon emissions reduction in China.

Further enhancements to the SSI nowcasting system should focus on the calibration of the SSI estimation model and improving its applicability for China and FY-4A. More specifically, the coefficients of the empirical method can be refined if high-quality SSI measurements are available across the country.

Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 42030608, 41805021, and 51776051), the Beijing Natural Science Foundation (Grant No. 8204072), and Beijing Nova Program (Grant No. Z211100002121077). We would like to thank the National Satellite Meteorological Center and the MERRA-2 teams for providing the data used in this study. The FY-4A data were collected from http://satellite.nsmc.org.cn/portalsite/default.aspx (accessed on 20 June 2021). The MERRA-2 data were collected from http://disc.sci.gsfc.nasa.gov/ (accessed on 12 September 2021).