A novel approach for the nowcasting of clouds and direct normal irradiance (DNI) based on the Spinning Enhanced Visible and Infrared Imager (SEVIRI) aboard the geostationary Meteosat Second Generation (MSG) satellite is presented for a forecast horizon up to 120 min. The basis of the algorithm is an optical flow method to derive cloud motion vectors for all cloudy pixels. To facilitate forecasts over a relevant time period, a classification of clouds into objects and a weighted triangular interpolation of clear-sky regions are used. Low and high level clouds are forecasted separately because they show different velocities and motion directions. Additionally a distinction in advective and convective clouds together with an intensity correction for quickly thinning convective clouds is integrated. The DNI is calculated from the forecasted optical thickness of the low and high level clouds. In order to quantitatively assess the performance of the algorithm, a forecast validation against MSG/SEVIRI observations is performed for a period of 2 months. Error rates and Hanssen–Kuiper skill scores are derived for forecasted cloud masks. For a forecast of 5 min for most cloud situations more than 95 % of all pixels are predicted correctly cloudy or clear. This number decreases to 80–95 % for a forecast of 2 h depending on cloud type and vertical cloud level. Hanssen–Kuiper skill scores for cloud mask go down to 0.6–0.7 for a 2 h forecast. Compared to persistence an improvement of forecast horizon by a factor of 2 is reached for all forecasts up to 2 h. A comparison of forecasted optical thickness distributions and DNI against observations yields correlation coefficients larger than 0.9 for 15 min forecasts and around 0.65 for 2 h forecasts.
Availability of energy power plays a central role in society and its economical evolution. Among the renewable energy sources, concentrating solar power (CSP) systems have a great potential since they combine electricity production with a storage capacity. By means of mirrors the incoming solar irradiance is concentrated, heating a fluid and driving a heat engine. The used technologies are parabolic trough, solar power tower, Fresnel reflectors and dish Stirling. In case of low insolation the electricity production is taken over by a fuel, e.g., gas. The operation of such solar power plants is challenging since the thermodynamic properties of the heated fluid are difficult to control, for instance when the CSP plant is only partially illuminated by the Sun or when insolation is strongly variable over time ranges of a few minutes to a few hours.
The fuel of solar power plants is direct normal irradiance (DNI). The main
source of DNI spatiotemporal
variability is cloudiness due to its intrinsic spatiotemporal
inhomogeneity and to the fact that already thin clouds can reduce DNI to
unusable levels for CSP. Further factors that affect DNI are aerosols and, to
a lesser extent, water vapor and ozone
Convective clouds are of particular interest for society due to the high
precipitation rates that are often connected to them. Because of their rapid
development they increase the error in any forecast. Therefore it is
reasonable to treat convective and advective clouds separately and to
investigate the development of convective cells, which is a challenging task.
During the last decades a large number of different cloud nowcasting
approaches have been developed, most of them with a strong focus on
thunderstorms. These techniques are using near real-time information from
radars, e.g., CONRAD
This publication presents a novel nowcasting algorithm based on satellite data from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) aboard Meteosat Second Generation (MSG) for all clouds with a focus on clouds which are relevant for CSP generation. With its high repetition rate of 15 min, its spatial sampling distance of 3 km and the availability of 12 spectral channels, this sensor is very well suited for the determination and forecast of cloud optical properties to be used to derive DNI since clouds are highly variable in space and time. Our method focuses on forecast times from 5 to 120 min. It exploits an optical flow algorithm to determine atmospheric motion vectors for every pixel. The starting point is represented by the optical thickness of clouds that are first split up into two (vertically overlapping) layers in order to take care of different velocities of upper level and low level clouds. To reduce the turbulent character of the atmospheric motion field on small scales, rendering long-range forecasts impossible, cloud subsets are defined as rigid objects that move with time. Convection cannot be forecasted adequately this way, but our approach considers dissipating convective clouds, where extended anvils are produced that can live for many hours and have an important impact on DNI. DNI itself is eventually computed from the cloud optical thickness forecast. A validation against MSG/SEVIRI observations is shown at the end.
After a description of the satellite instrument MSG/SEVIRI and the cloud
detection the analysis algorithms needed for our forecast method including
the optical flow procedure are presented (Sect.
MSG is a series of European geostationary
satellites operated by EUMETSAT. Their primary mission is the observation of
weather phenomena on the Earth's full disk with the SEVIRI imager. It consists of three channels in the
visible, one in the near infrared and eight in the infrared spectral range
with a sampling distance of 3 km at the sub-satellite point
Optical, micro- and macrophysical properties of clouds are
important parameters for the modeling of radiation–cloud interactions. Thus,
their determination plays an important role for the computation of surface
radiation, including DNI. The cloud detection and analysis algorithms used in
this work are presented in the following two sections (Sect.
For thin ice clouds the “Cirrus Optical properties derived from CALIOP and
SEVIRI during day and night” (COCS;
In addition to COCS, APICS (Algorithm for the Physical Investigation of
Clouds with SEVIRI;
A cloud top phase mask is produced by merging the results of the two algorithms. A cloud detected by COCS is an ice cloud, i.e., its cloud top is composed of ice. Clouds detected by APICS can be both ice and liquid water clouds. Thus, we assign all clouds detected by COCS to the ice phase, while a cloud that is detected by APICS but not by COCS is assigned to the liquid water phase.
Cloud optical properties are obtained the following way. COCS provides the
optical thickness of the upmost cloud layer. Furthermore, for both cloud
types (liquid water and ice), cloud optical thickness and effective radius
are derived from two solar channels using APICS. In APICS, SEVIRI channels
centered at 0.6 and 1.6
For the detection of convective clouds, procedures of the Cb-TRAM algorithm
(CumulonimBus TRacking And Monitoring;
convection initiation (stage 1) rapid cooling (stage 2) mature thunderstorm cells (stage 3).
In this work we consider only stage 3, the detection of mature thunderstorm
cells. It is limited to areas with a strong spatial roughness of the HRV,
determined by the local standard deviation, combined with the brightness
temperature difference of 6.2 and 10.8
The forecast rests upon an optical flow method determining a motion vector
field from two consecutive images which is part of Cb-TRAM
(
Movements in the atmosphere take place on different scales reaching from microscale (few centimeters) to global scale (10 000 km). These large-scale flows overlay the small-scale movements so that the determination of the disparity vector field for all scales is challenging. In order to take this into account the disparity vector fields are successively derived on different scales, starting from low resolution down to high resolution – a pyramidal scheme.
The procedure is described by means of an example for two images Select the number of sub-sampling levels Define the images Start the iterative process. Calculate the dimensions Resample the start images Determine comparison images Identify the best fit between all possible
To mitigate the impact of singular incorrect motion
derivations and ensure physically realistic local flow fields, these
initially integer displacements Blow up the resolution of Add the motion vectors obtained so far to
Warp the image for every pixel position Reduce the value of At the end of the iterative procedure, the refined disparity vector
field
The refined disparity vector field
For more details, technicalities and an additional example please see
In this section the forecast algorithm is described. It exploits the methods introduced in the previous section. First, a more advanced cloud classification is implemented that distinguishes two overlapping classes of clouds. Then, the pixel-based disparity vector field is determined for both cloud classes separately. Cloud objects are formed, based on optical thickness, and motion vectors are derived for these objects. After the assignment of motion vectors to cloud-free areas, clouds are warped to their new position with this motion vector field. An intensity correction is applied for rapidly thinning convective clouds. In a last step the DNI is calculated from the optical thickness.
In the following, clouds are classified in SEVIRI images according to two
criteria: The first one considers the cloud top phase and the vertical
structure of clouds (Sect.
Low level and high level clouds are often observed to move in different
directions at different velocities due to complex wind profiles in the
atmosphere. In order to take this aspect into account, we aim at the
separation of low and high level clouds and the generation of two forecasts,
one for low level and one for high level clouds. However, using APICS and
COCS applied to SEVIRI satellite data according to Sect.
In general, the discrimination among all these cases and the determination of
optical properties for all cloud layers is challenging using only passive
satellite observations. Several approaches have been proposed
Qualitative indications contained in SEVIRI's spectral channels can be
exploited to provide a reasonable differentiation between one layer and two
layer cloud situations. For clarification an example is depicted in
Fig.
Assignment of cloud optical thickness to
two cloud classes called upper clouds and lower clouds.
Liquid water clouds identified following Sect.
The results of the classification applied to Fig.
The focus of the presented forecast method is the accurate
prediction of thin ice clouds since they modulate surface DNI in the relevant
range for CSP. Often ice clouds are formed by convection. In contrast to most
ice clouds that are mainly characterized by horizontal advection, convective
clouds show a strong local vertical development. While during growth and
maturity of convective cells large optical thickness values dominate and DNI
at surface is negligible, anvil ice clouds formed during maturity can live
much longer than the thunderstorm cloud itself during the decaying stage
For this reason, a third class of clouds is defined: we single out mature
convective clouds using the stage 3 detection of the Cb-TRAM algorithm as
discussed in Sect.
Once clouds have been classified and cloud optical thickness has been
determined (Sect.
In this first stage motion vector fields are derived for the optical
thickness of lower clouds and upper clouds separately. Since convective
clouds as defined in Sect.
There are two reasons for the use of the optical thickness as input parameter
for the optical flow method: first, it is the quantity need for the
calculation of DNI (see Sect.
Forecasts are produced in forecast steps of
Illustration of the forecast of optical thickness
for upper clouds for 7 April 2013
To illustrate the result of this forecast procedure we consider the upper
cloud layer from the example in Fig.
Upper cloud layer optical thickness extracted
from the lower left part of Fig.
For the reason discussed above, an averaging procedure for the pixel-based cloud motion vectors is implemented. To this end, neighboring pixels with similar optical thickness are combined to objects. This procedure is called object classification and is applied separately to upper and lower cloud layers since they are forecasted separately. At this step, convective clouds are treated separately. This averaging procedure removes small-scale variability that is realistic at the moment of derivation but makes the forecast unstable.
For upper clouds, first each convective cell (Sect.
For the lower clouds the object classification is performed in a similar way:
optical thickness in the range
Next, a mean motion vector is calculated for each object and this vector is
assigned to every pixel in the object; i.e., the object
moves as a whole during the
forecasting procedure. The forecast image produced this way is called
object-based forecast. An example is shown in Fig.
Delaunay triangulation for the
As the motion vectors are derived from cloud optical thickness, the disparity
vector field in the area between the clouds goes to zero
(Fig.
Domain used for the classification of decaying cells
and for the validation presented in Sect.
The pyramidal matcher (Sect.
Distribution of the change in ice optical thickness in relation to the divergence with blue crosses denoting the decaying cells and red for the non-decaying cells.
Distribution of the decrease factor
To this end, quickly thinning convective clouds are first identified in
satellite data and the successive evolution of their optical thickness, as
far as it can be forecasted through disparity vectors, is then corrected to
follow typical temporal patterns. Both the identification of these clouds and
the determination of typical values for the temporal evolution have been
developed based on 300 cells detected by Cb-TRAM (stage 3, mature cells,
according to the classification presented in Sect. mean change in optical thickness from one time to the next is
smaller than mean divergence div (Eq.
Thus, a decrease in optical thickness and a slight converging movement
indicate a decaying cell.
Upper cloud optical thickness for 9 June 2013, 17:00 UTC, for the real situation (middle) compared to the object-based forecast for 30 min (left) and the forecast with intensity correction (right).
To determine a typical correction term for the temporal evolution of upper
cloud optical thickness after the decaying phase has started, the subset of
all 70 decaying cells has been investigated closer. An empirical modification
derived from them is imposed onto the optical thickness of the convective
objects forecasted through the disparity vectors. Before the application of
disparity vectors as described in Sect.
One example of a decaying cell is shown in Fig.
After classification into upper and lower layer, object-based forecast of
these layers and special correction for decaying upper layer convective
cirrus clouds, two fields of possibly overlapping forecast optical thickness
are available. Figure
Optical thickness for upper and lower clouds together
for 7 April 2013, 13:15 UTC (left), and the calculated direct normal
irradiance in W m
DNI computed in this paper considers only photons coming from the Sun that do
not interact with the atmosphere (see the “strict definition” of DNI for
numerical modeling of radiative transfer in
Figure
For a validation of cloud and DNI forecasts two time periods,
4–31 March 2013 and 12–31 July 2013, were examined. These 2 months were
chosen due to the appearance of different cloud types. The domain considered
is the central part (marked in red in Fig.
Contingency table.
In order to quantitatively assess the performance of the forecast algorithm,
we evaluate its capability to predict clouds and cloud-free pixels by
examining the errors of the forecast cloud mask against observed cloud masks
from SEVIRI. Observations and forecast are connected through the contingency
table (Table
Illustration of the elements of the contingency table for upper (left) and lower (right) clouds with regard to the 1 h forecast for 7 April 2013, 14:15 UTC: hits in red, false alarms (fa) in blue, misses in green and correct negatives (cn) in white.
The calculated parameters of the contingency table for all start times are
averaged for every forecast time step up to 2 h (see beginning of
Sect. the difficulty
of a retrieval for lower clouds below thick upper clouds, which leads to
errors in the forecast when not all clouds are detected in the initial
images, a low cloud layer disappears below a high one or low clouds evolve
into high clouds; larger small-scale variability for lower clouds, which
cannot be resolved by SEVIRI sub-pixel
inhomogeneity, i.e., broken cumulus cloud fields and rapidly changing
small-scale convective cloud fields with very short timescales but low
advection speeds formation of new lower clouds that
cannot be forecasted.
Errors are smaller in July compared to March because of the low cloud cover
of 22.1 % on average (62.7 % in March) during this month (errors are
relative to satellite scene size). For water clouds, detection and forecast
are hindered by the presence of upper clouds such that even a correct
forecast might be incorrectly classified. Thus, it is difficult to assess the
real accuracy of water cloud forecasts. For this reason, persistence for
water clouds has not been evaluated.
In addition to the evaluation of the errors as shown above, we apply the
Hanssen–Kuiper skill score
The HK, also called Hanssen–Kuiper discriminant, Peirce skill score
Two-dimensional histogram of the forecasted optical
thickness of the upper cloud layer compared to the real optical thickness
with forecasts starting at 13:00 UTC every day: for a 15 min forecast
Two-dimensional histogram of the forecasted optical thickness
for the lower cloud layer compared to the real optical thickness with
forecasts starting at 13:00 UTC every day: for a
15 min forecast
The resulting HK (Fig.
Two-dimensional histogram of the observed optical
thickness of the upper cloud layer compared to the persistence optical
thickness for a time difference of 15 min
Two-dimensional histogram of the forecasted DNI compared to the
observed DNI with forecasts starting at 13:00 UTC every day: for a
15 min forecast
In order to test the performance of the algorithm with regard to the optical
thickness a comparison of the forecasted optical thickness with the optical
thickness observed from SEVIRI is done via a 2-D histogram
separated into upper (Fig.
For the upper cloud layer the algorithm shows an overall good performance
with only small differences for most of the pixels for the 15 min forecast
(Fig.
For the lower cloud layer (Fig.
The corresponding correlation coefficients show high values above 0.74 for the first time steps despite the lack in skill for cloud detection of lower clouds. Its deteriorating influence is apparent in the sharp decrease of the correlation coefficients for the 1 and 2 h forecast.
To judge the quality of the forecast algorithm, histograms of the persistence
method for the upper cloud layer are shown in Fig.
Figure
The correlation coefficients for DNI are mostly higher than the values of both cloud types, especially for long forecasts, with lower values in July. This is due to the fact that forecasts are better for cloud areas with small optical thickness values than for optically thick clouds. Derivation of DNI emphasizes the relevance of these thin clouds for DNI predictions, while errors in the forecast of thick opaque clouds (e.g., new convective developments) are less detrimental.
Based on an optical flow method deriving cloud motion between two consecutive images, an algorithm for the forecast of cloud optical thickness and DNI has been developed for input data from the imager SEVIRI aboard the geostationary MSG satellite. The algorithms COCS and APICS provide cloud detection and cloud optical thickness for two vertically separated layers. Because of different velocities and motion directions these low and high level clouds are forecasted separately for a time step of 5 min with a forecast horizon up to 2 h. To deal with the small-scale fluctuations of the motion field derived for the two levels, which would spoil forecasts of more than 15 or 30 min, an object classification is applied to the cloud layers and cloud-free background motion is interpolated. An intensity correction for decaying convective cells is implemented.
Using the satellite observations, the forecasts for March and July 2013 have been quantitatively validated. As far as cloud detection is concerned, the largest inaccuracy consists in the difficulty to retrieve clouds below optically thick ice clouds. Consequently, forecast errors for the lower cloud layer are considerably higher than for high clouds. The forecast accuracy also differs for the two time periods because of different cloud coverage and cloud types. In March mainly fronts with many advective multilayer clouds dominate in contrast to a high amount of shorter-lived convective clouds in July.
Convective clouds during July cause the forecast skill to decay quicker with forecast horizon than during March. For any given forecast quality requirement, over all cloud (or weather) types and for both cloud layers, a doubling of lead time is found comparing the developed forecast to a non-forecast, i.e., persistence.
The impact of weather situation also becomes apparent in the comparison of
observed and forecasted optical thickness. The distribution of deviations,
analyzed by means of 2-D histograms, as well as the correlation between
forecast and observation show better results for March, especially for a
longer forecast. The wider scatter of deviations as well as lower correlation
coefficients confirm the limitations of the forecast quality for low clouds
compared to high clouds. Although much effort was invested in the
identification of multilayer clouds and their differential motion, this
still remains a main source of uncertainty for satellite-based nowcasting. An
additional source of uncertainty is sub-pixel inhomogeneity, i.e., broken
cumulus cloud fields and rapidly changing small-scale convective cloud fields
with very short timescales but low advection speeds
Finally, DNI forecast verification shows that most correct forecasts are, of course, found for the expected clear-sky DNI and no direct irradiance below thick clouds. However, the overall correlation between the 2 h forecast and the observation is still around 0.7.
As a next step comparisons to ground-based irradiance measurements shall be conducted. To this end, the DNI model should be extended to consider varying trace gas concentrations and aerosol loads, e.g., water vapor from numerical weather models and aerosol information from ground-based networks (e.g., AERONET).
An extension of the forecast horizon up to 5–6 h could be performed, which
is indeed useful for CSP operators. For longer forecast horizons forecasting
methods of NWP models have a higher accuracy compared to satellite-based
methods
The MSG/SEVIRI level 1.5 data are available at
The authors declare that they have no conflict of interest.
We acknowledge the European Commission for funding the project DNICast
(