Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Applied Energy 206 (2017) 1464–1483

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Short-term peer-to-peer solar forecasting in a network of photovoltaic T


systems

Boudewijn Elsinga , Wilfried G.J.H.M. van Sark
Utrecht University, Copernicus Institute of Sustainable Development, PO Box 80.115, 3508 TC Utrecht, The Netherlands

H I G H L I G H T S

• Network of photovoltaic systems as input for clearness index time series.


• Time lag between locations to obtain < 30 min forecasts at 60 s resolution.
• Forecast Skill improves with daily irradiance variability.
• Averaged over 11 locations over 56 highly variable days, Forecast Skill of 5.99% obtained.

A R T I C L E I N F O A B S T R A C T

Keywords: Solar forecasting is a necessary component of economical realization of high penetration levels of photovoltaic
Solar forecasting (PV) systems. This paper presents a short term, intra-hour solar forecasting method. This “peer-to-peer” (P2P)
Intra-hour forecasting method is based on the cross-correlation time lag between clear-sky index time series of pairs of PV-
Sensor network systems that are influenced by the (assumed) same cloud sequentially, with the feature that the forecast horizon
Time lag correlation
(FH) can be set at a fixed value. The P2P forecasting algorithm was evaluated for 11 central PV-systems (out of
Irradiance variability
202) over a half year period from the 1st of March through the 31st of August 2015 using the forecast skill (FS)
metric. Positive FS means improvement over reference clear-sky index persistence forecasting. The P2P fore-
casting method was evaluated over a subset of days with either high, all or low irradiance variability. The
average forecast skill (avgFS) concerning forecast horizons between 5 and 8 min was 5.99%, −1.61% and
−16.0% over these periods respectively, indicating the superior performance of the P2P method over persis-
tence during the highly variable days, which are most interesting from the perspective of electricity grid
management.

1. Introduction of reliable, high resolution measurement data and processing cap-


abilities.
With the rapid and continuing increase of intermittent solar pho- Ad hoc curtailment or fixed capping of PV production (which can be
tovoltaics (PV) electricity generation capacity, surpassing a global ca- seen as wasting free energy), demand side management (DSM) by
pacity of 300 GWp by the end of 2016 [1], the intermittency in the shifting planned loads; and on the long term, (costly) grid reinforce-
electricity grid needs to be addressed to ensure reliable and proper ment (also on the MV and LV levels) can be regarded as counter-
functioning of that grid [2]. It is therefore beneficial to the operation measures to cope with issues (power flow bottlenecks, reduced power
and safety of the electricity grid to have knowledge of the amount of quality, etc.) on the distribution level [8]. As a compromise, planning of
expected produced power and its variability, on multiple spatial and tilt and orientation of newly or (re)designed PV-systems can be such
temporal scales: whether it is for managing reserve capacity of con- that the peak during solar noon (when the irradiance on south-facing
ventional, central electricity generation hours or day-ahead or to bal- PV panels is maximum) is shifted to the morning or evening by or-
ance electricity supply and demand on intra-hour time scales in a future ienting the panels to the East and/or West, respectively [9]. Apart from
smart grid [3–7]. As a result, the field of solar PV forecasting has seen flattening the cumulative PV electricity generation in an area, the di-
increasing interest among researchers around the world, partly driven rectly consumed part of the generated electricity (self-consumption) is
by the increase of installed PV capacity, but also due to the availability higher as production is shifted more towards the morning and evening


Corresponding author.
E-mail address: b.elsinga@uu.nl (B. Elsinga).

http://dx.doi.org/10.1016/j.apenergy.2017.09.115
Received 9 June 2017; Received in revised form 20 September 2017; Accepted 21 September 2017
Available online 30 September 2017
0306-2619/ © 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/).
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

rush hours, which can have financial benefits if the excess electricity is deterministic forecasting. The former relies on ensemble data and
sold at a lower price than the consumer retail price, known and/or assumed uncertainty and produces forecasts along with
Extensive grid reinforcement might be implemented less intrusive or uncertainty ranges. The latter relies either on the time series of one PV-
avoided partly if the mismatch of supply and demand of electricity can system and/or multiple time series and usually in combination with a
be guaranteed to be flattened, through smarter management of the physics-based model that relates the input data to a single forecast
generators and consumers of electricity in the grid, including smart value. Within both of these categories, the used input data can vary
(reactive) power control by modern inverters with or without local from using only locally measured power output and/or including his-
battery storage [10]. In general, Demand Response (DR) or Demand torical data and/or including multiple locations and/or including exo-
Side Management (DSM) is seen as a option to this effect, albeit that the genous data (like measured weather variables as wind vector, cloud
options are limited as the use of electric appliances is largely ad hoc and cover, temperature, etc.). A special category of inputs are covered by
wet appliances (dishwasher, washing machine) and to some extend re- satellite image or remote sensing driven forecasts [22,17], that can be
frigerators can be seen as plannable loads leading to 1.5–4% increase of (partially) supplemented by numerical weather prediction (NWP)
directly consumed PV electricity that otherwise would have been in- models that find their roots in meteorology and broader weather fore-
jected into the grid for a sample of 100 Dutch households [11]. casting, see e.g. Pelland et al. [23], Mathiesen et al. [24], Mathiesen
There is no single blueprint for a smart grid that couples energy and and Kleissl [25], Perez et al. [26]. As an alternative to satellite images,
data flows in order to facilitate the growing asymmetry in a once one- ground-based photographs of the sky (using fish-eye cameras or curved
directional electricity market and grid at the distribution and trans- mirrors) can be used to track local clouds and predict the (direct) ir-
mission level [12]. In any form, ideal operation of local electricity radiance for very short range forecasts (minutes ahead), see re-
storage needs knowledge of future (forecasted) demand and supply of presentative examples by Marquez and Coimbra [27], Chow et al.
electricity in order to optimize the charging strategy of the energy [28,29], Chu et al. [30,31], Peng et al. [32].
storage system that are bound by requirements for economics, state-of- In terms of purely time series driven forecast methods, various le-
charge, charging cycles, self-consumption and reserve capacity to name vels of autoregressive methods, that are rooted in econometric fore-
a few. Short-term electricity storage can serve many non-excluding casting, have been pursued. Simple Markov-Chain forecasts based on
goals: peak-shaving and buffering for grid stability and increased self- conditional (trained) probability between discrete irradiance classes are
consumption of generated electricity, optimizing economic value of reported by e.g. Muselli et al. [33], Li et al. [34]. The k-nearest
generated PV under variable market price, or island operation of a neighbors (k-NN) algorithm that produces forecasts based on agnostic
house or neighborhood in remote areas [13]. historic similarity of time series is described by Duda & Hart [35] and
In the Netherlands, a Groningen neighborhood with lots of PV and a e.g. by Yakowitz [36]. A large class of methods that combine auto-re-
pilot smart grid Power Matching City [14] and in the Utrecht district of gression (AR) and/or moving averages (MA) of historic data is called
Lombok the first European bidirectional solar charged electric vehicle ARIMA (Auto Regressive Integrated Moving Average) and is used by
(EV) charging station was opened in 2016 as part of a local PV-EV real e.g. Reikard [37], Bacher et al. [38], Yang et al. [39] for solar fore-
life testing environment [15] are examples of smart grid pilot projects casting. For general details on ARIMA methods, see Hyndman et al.
in the Netherlands. Synergy between electric vehicles and VRE (by [40].
vehicle-to-grid V2G) is thought of as a major contribution to expanding Advanced non-linear methods of data processing, in which an (in
the penetration of VRE on the LV grid, especially PV [16,3]. general) agnostic computer model is trained on historic data, include
Solar forecasting deals with the prediction of PV power production random forest (RF) search methods [41], the training of artificial neural
(or irradiance as a proxy for power). Depending on the temporal scope, networks (ANN) by e.g. Sfetsos and Coonick [42], Paoli et al. [43],
different families of methods exist that use different data inputs and Voyant et al. [44], Pedro and Coimbra [45] and the use of support
forecast for different time horizons and at different levels of detail, see vector regression (SVR) as can be seen in work by Fonecsa et al. [46],
Table 1. To each of the presented temporal scales, a set of targets can be Boland [47], Rana et al. [48]. The general term for all these forecasting
associated in relation to what solar forecasting is meant to improve or methods is Machine Learning (ML). Additionally, regressive, numerical
assist in the context of the electricity grid [17–20]. and/or ML forecasting methods can be merged into hybrid methods and
exogenous (weather) variables may be included to improve forecast
accuracy, see some examples of this type in work by: Marquez and
1.1. Overview of solar forecast methods Coimbra [27], Dambreville et al. [49], Zagouras et al. [50], Wolff et al.
[51], Vaz et al. [52]. It is not our intention to give a detailed overview
Forecasting methods can be subdivided in two categories, according or review of all of the available forecast methods, we rather refer to
to Antonanzas et al. [21]: (1) probabilistic forecasting, and (2) recent detailed reviews on solar energy forecast methods, such as by:
Inman et al. [53], Widén et al. [54], Antonanzas et al. [21], Barbieri
Table 1 et al. [55] or more specifically Machine Learing by: Mellit and Kalo-
Overview of different temporal variability and forecast horizon scales and the associated
girou [56], Voyant et al. [57], or more general overviews of solar
purposes (target) and forecast method families.
forecasting in: Lorenz and Heinemann [17], Kleissl [19]. Sensor net-
Time horizon < 15 min 15 min–h h–day > day work based time series forecasting methods that do not rely heavily on
AR or ML will be discussed separately in Section 1.3.
Target
Power balance/quality •
Reserve capacity planning • •
1.1.1. Correlation and smoothing
Load following • • Many of the aforementioned forecasting methods process and pre-
Market bidding • • • dict irradiance or power on hourly time scales. This resolution is suited
Base-load planning • • for day-ahead energy production forecasting and/or for large spatial
aggregation (many PV-systems in an ensemble or a large PV plant).
Forecast method family
However, at lower spatial aggregation (individual PV-systems) and at
Numerical Weather Prediction (NWP) ∘ ∘ ∘ higher (sub-hourly) temporal resolution, variability is smoothened
Satellite data/images ∘ ∘ much less and the details of local cloud shades becomes more im-
All-sky imaging ∘ ∘
portant. Correlation of variability has found to be inversely dependent
Time series, statistical ∘ ∘ ∘
Time series, network ∘ ∘ on inter-system distance, resulting in smoothing of aggregated power
production over large areas (hundreds of kilometers; hours), described

1465
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

by e.g. Murata and Otani [58], Wiemken et al. [59], Perez et al. [60], 1.3.3. Peak Matching Algorithms
Hoff and Perez [61], Perez and Fthenakis [62], or on smaller intra-hour In order to circumvent accurately measuring and tracking elusive
time scales within dense sensor networks (kilometer scale) Lave et al. cloud edges, Achleitner et al. [76] describe a method (SIPS, solar irra-
[63], Hinkelman [64], Perpiñán et al. [65], Elsinga and Van Sark [66], diance prediction system) that forecasts power output of a PV-farm
and Lohmann et al. [67]. (“solar field”) that is approx. one kilometer downwind of an array of
irradiance sensors, completed by a sky-imager. They introduce the Peak
Matching Algorithm (PMA) that matches peak values of measured data
1.2. Problem definition
at the sensors and the PV-farm to establish the momentary time lag in
the cloud shadows moving from the sensors in the direction of the PV-
Next to forecasting methods that either rely on well-studied, but
farm. This time lag is then used to predict the output of the PV-farm in
coarse forecasting methods like NWP and/or satellite or finer resolution
the near future, equal to the found time lag by shifting the previous
data-driven statistical/black-box methods in which the forecasting is
measurements of the best matching sensor by this time lag ahead. This
based on historical resemblance/regression, we want to introduce a
method, using 5 s input data is very successful at forecasting at very
lean short-term deterministic time series forecasting algorithm that is
short time scales (up to 75 s), but is limited to the cloud motion vectors
based on the local measurement in a dense monitored PV-system net-
in the direction between the sensors and the PV-farm. Moreover it
work (the PV sensor network) without reference to or training on
produces forecasts of variable forecast horizon. Lipperheide et al. [77]
(autoregressive) historic data beyond the present local cloud field: we
approach the forecasting of very short forecast horizon (nowcasting) of
denote this as the peer-to-peer (P2P) forecasting algorithm (based on
sub-inverters of a large PV-farm by determining the cloud velocity
preliminary work, reported by Elsinga and Van Sark [68]). The P2P
vector from triangulation of triplets of reference solar cells and a PMA
forecasting algorithm or P2P method is suited for the forecast of global
that is described in [78,79]. Forecasts are generated by transposing the
horizontal irradiance (GHI) of rooftop PV-systems that forecasts several
measured power at all of the sub-inverters (that act as proxies for ir-
minutes ahead with high (several seconds to several minutes) temporal
radiance sensors as all the PV-modules are identically set-up) in the
resolution.
direction of the found cloud motion vector. Parts of the PV-farm that
are upwind are forecast by ramp persistence. The (variable) forecast
1.3. Comparable forecasting methods horizon is limited to < 180 s due to the limited spatial dimensions of
the used part of the PV-farm (1807 m × 539 m). The data resolution is
It is our aim to investigate in what way we can improve on/extend 1 s.
previous methods that operate in approximately the same regime.
1.4. Aim of the paper
1.3.1. Cloud advection forecasting
The main methodological improvement of our P2P method over the
Lonij et al. [69] use four methods to estimate the cloud motion
ones mentioned above (especially the PMA-related ones), is the possi-
vector to transpose the measured power (normalized to clear sky
bility to choose a fixed forecast horizon much larger than the data
power) of 80 rooftop PV-systems in Tuscon, Arizona. Forecasts for
temporal resolution as with traditional one-step-ahead forecasting.
target locations are then based on the median of the transposed mea-
Furthermore, the large spatial extent (≈1400 km2) and approximately
surements at the closest PV-systems. Forecast horizons in steps of
concentric lay-out of the PV-systems in our network ensures full flex-
15 min are reported up to 90 min ahead. The measurement data is of
ibility in coping with variable cloud motion vectors. Implicitly, we also
15 min temporal resolution. They conclude their paper by describing
demonstrate the use of monitored PV-systems as GHI sensors by de-
the correlation of the shifted measurements to estimate the accuracy of
ploying an Inverse PV model [80] and study the relation of the fore-
the forecasting. Lorenzo et al. [70] use a network of GHI sensors de-
casting method over a half year measuring period in relation to the
ployed upwind of a neighborhood (also Tuscon, Arizona) where PV-
daily irradiance characteristics. Eventually, forecasting of GHI is im-
systems are monitored, with respect to the local prevailing wind di-
portant for the forecast of intermittent PV power for generic PV-mod-
rection. Forecasts are made based on transposition of the interpolated
ules in order to optimize short-term DSM control strategies (EV (dis)
irradiance field with a wind vector that is measured at maximum hu-
charging, pre-heating or -cooling, etc.) in local smart electricity grids.
midity height (corresponding to cloud height) and compared to naïve
spatial and temporal averaging of the measured values. Their study,
2. System description and data treatment
that builds on the work of Lonij et al. [69] to some extent, also in-
dicated that the inclusion of sensors upwind has a strong positive effect
202 rooftop PV-systems are connected with commercial off-the-shelf
on the error metrics of the forecast method. The data temporal re-
but calibrated Upp-energy power measurement data loggers [81] in the
solution (1 s) and forecast horizon (several (tens of) minutes ahead) are
Province of Utrecht, the Netherlands; this is called the “PV-Sensor
comparable to the P2P method described in the present paper.
Field”, see Fig. 1. The power data is recorded at a resolution of 0.7 W
and stored real time at approx. 2 s interval in a database. The power
1.3.2. Spatio-temporal correlations time series are then converted into Global Horizontal Irradiance (GHI)
Yang et al. [71] implement the lasso parameter shrinkage method to time series, by correcting for the tilt and orientation, inverter power
select highly correlating lagged neighbors (and/or the autocorrelated and (hourly) interpolated ambient temperature using an inverse PV-
time series of a target location) in a dense GHI sensor network. Sta- irradiance model [80]; this inverse PV model is based on inversion of a
tionary spatio-temporal correlation is inferred by training on recent modified version of the decomposition model of Orgill and Hollands
measurement history and the notion of “up wind” and “downwind” is [82] and the diffuse irradiance transposition model developed by Perez
assumed static (which may be the case for the studied period (13 days) et al. [83].
and location (Oahu, Hawaï), but is not easily generalized; however, the The GHI time series is subsequently converted into terrestrial clear-
authors also provide discussion on their method in absence of wind sky index, kt∗, by dividing the GHI by modeled clear sky GHI, obtained
knowledge.). Their forecast model performs very well for short time from the Ineichen clear sky model [85]. This way, a perfect and shade
scales (< 5 min) in a dense (1.2 km × 1 km) network of 17 GHI sen- free PV-system would record a clear, sunny day as a constant kt∗ = 1 all
sors. In related work, [72–74], this spatio-temporal methodology is day. Deviations from this value are interpreted as shading/super-irra-
extended to unobserved locations using kriging; a weighted interpola- diance from clouds. Assumed that the system data (tilt, orientation,
tion technique originating in geo-statistics [75]. installed capacity, etc.) of the PV-systems, mostly provided by the PV-

1466
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 1. Left Overview of the “PV-Sensor Field”. Shown are the locations where an Upp data logger is installed; the smaller dots are locations left out of the study due to inverter under-
dimensioning. This map shows the Province of Utrecht (highlighted area of approx 1400 km2) and the clusters of points show roughly where the population centers are, most notably the
city of Utrecht in the center. Highlighted locations in the center of the map (red) indicate the PV-systems that were used for generating results. Right This graph shows the (unique) inter-
system distance histogram of the shown locations. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Map data: Wolfram Research Inc. [84]/OpenStreetMap.

system owners themselves, correspond with the actual situation at the soiling or other systematic errors, an overall correction factor was ap-
locations, the resulting kt∗ time series can be used directly to predict kt∗ plied on each of the PV-systems GHI data. This factor was found by
at a chosen PV-system in the PV-Sensor Field. In other words, clear sky- comparing GHI data to modeled clear sky GHI data on days with per-
index is independent of tilt, orientation and time of the day. The PV- iods of clear sky. A simple linear regression (forced through 0) was
systems that constitute the PV-Sensor Field are selected to have as little performed on the scattered GHI(t) versus clear sky GHI per PV-system.
surrounding-related (topological) shade as possible, but a trade-off has The section of day where the regression was found with the highest
been made between distribution of PV-systems and shading to let the R2 -value was taken as the inverse of a factor by which the time series of
spatial distribution be as homogeneous as possible. Forecasting of clear- the entire day of that PV-system is to be multiplied. This way, the re-
sky index is naturally suited for applying clear-sky index persistence gression is impervious for small shading events or even topological
and is a means to eventually forecast GHI. shading during an otherwise clear day. Finally, the mean of correction
The data used for the numerical validation of our method is down- factors obtained for 6 clear days throughout the year was applied per
sampled by taking the mean value over bins of Δt = 60 s. Missing values PV-system, these factors are in the order of a few percent.
were interpolated at 1st order. PV-systems that have an outage of more
than one hour during a day are also omitted during that day.
Additionally, to increase the reliability of using PV-systems as GHI 3. Peer-to-peer forecasting method
sensors, PV-systems that have an inverter of rated capacity [W] lower
than the rated capacity of the solar panels [Wp], were omitted: this In principle, the P2P forecasting method is straightforward: clouds
clipping effect obscures the output beyond the maximum (rated) in- that move over the PV Sensor Field will show as dips in the kt∗ time
verter power. This measure was necessary to ensure that all irradiance series of the PV-systems and influence several PV-systems sequentially.
variability is visible through registered PV output power which is es- A PV-system that receives the shadow first, serves as an indicator of
sential to the pattern-correlation in the Selection Algorithm. From these shading for PV-systems downwind. The forecast is then based on the
133 remaining locations, 11 PV-systems centrally located in the delay time, or time lag between an upwind PV-system and the target PV-
Lombok district of Utrecht city were selected to act as target systems for system. In our method, the generated forecast will be constructed by
generating the results, see Fig. 1. The central location of these PV-sys- successive influence of clouds that are actually and currently influen-
tems ensures the existence of sufficient peripheral PV-systems most of cing a PV-system, rather than very similar, but historic clouds, as with
the time. e.g. the k-NN method. It is assumed that the shape and speed of the
Forecasts are made based on a time range from three hours after clouds remains approximately constant during the transit; the P2P
sunrise through two hours before sunset on a given day, to minimize the method actively checks this assumption by measuring the quality of the
effect of topological shading that might occur on the PV-systems at low found correlation between PV-systems, this is explained in the next
elevation angles, especially in the built environment. Sunrise and sunset sections.
times calculations as well as the development and testing of the fore- An overview of the processing of the input data in the P2P method is
casting algorithm was done with Mathematica 11 Wolfram Research shown in Fig. 2. The individual steps will be treated in the subsequent
Inc. [84]. paragraphs.
Furthermore, to minimize the error from PV-system data, aging,

1467
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 2. Diagram showing the flow of the complete P2P forecasting algo-
Power measurements rithm. From top to bottom this diagram describes the steps of processing
(at PV-systems) input data time series, selecting peripheral locations and computing the
forecast time series for a given target PV-system i at time t0 and a chosen
Forecast Horizon (FH). The numbers between parentheses indicate the
System and data (2)
sections that describe the processing steps.

Clear-sky index
time series
For each target PV-system

Selection (3.1)
based on cross-correlation
For each moment

Peripheral (lagged)
clear-sky index ZZ
time series
Forecast (3.2)
at chosen FH

Forecast clear-sky index and


GHI time series

Error metrics Results (3.4 & 4) Aggregate


for PV-system Results

3.1. Selection algorithm window, TObs (t ,Δt ,Obs) and the variable δt by which the k ∗j,t time series is
shifted relative to ki∗,t . The peripheral time series k ∗j,t is shifted forward in
To be able to make a forecast of the clear-sky index kt∗ it is necessary time with respect to ki∗,t by varying δt and the resulting Y is evaluated at
to determine which PV-systems have time series that are suitable for each δt , see Fig. 3(c) for a visualization. There are multiple options for
forecasting the target PV-system. the indicator function: it must fulfill the requirement that it calculates
the overlap or sameness of two time series. The absolute value of this
3.1.1. Indicator function and time lag indicator function Y is not relevant, but rather the location of the
We consider two PV-systems at different locations: i and j. Let maximum, if there is any. The higher the value of Y, the better the
system i be the target system, and j be the peripheral system. The kt∗ of overlap of the ki∗,t and k ∗j,t − δt time series within the observation window.
the two PV-systems will be compared within the observation window, Correlation of the overlap between ki∗,t and k ∗j,t − δt , rewards similarity
TObs (t ,Δt ,Obs) which is a time window of length Δt,Obs in n steps of the in the trend in those time series, using Pearson’s correlation coefficient ρ ,
measurement time interval Δt until t, see Eq. (1) and Fig. 3. TObs reaches which essentially is the Covariance, normalized by the square root
back 60 min so the comparison of time series starts effectively from two product of the respective Variances, see Eq. (2).
hours after sunrise. Cov(TX ,TY )
ρ (TX ,TY ) = ,
TObs (t ,Δt ,Obs) ≡ {t −Δt ,Obs,…,t −Δt ,t }, Var(TX ) Var(TY ) (2)
where Δt ,Obs ≡ nΔt (1)
with TX and TY time series of equal length and nonzero Variance. This
Whether and to what extent the kt∗ time series of systems i and j match, results in the definition of the indicator function Yij , see Eq. (3). Yij
is determined by an indicator function, Yij (δt ) . This is a function of the ranges from −1 and 1 and has a maximum at perfect correlation when
kt∗ time series of the PV-systems at the locations i and j, the observation ρ = 1.

Fig. 3. Two close PV-systems (6.7 km apart) that have similar, lagged kt∗ time series due to passing clouds on June, 3rd 2015 between 7:14 and 9:44 UTC. Comparison is done by shifting
the red time series (k∗j (t ) ) forward by δt with steps of Δt = 1 min: see sub-figures (a) and (b). Resulting correlation as a function of δt over the observation window (between the vertical
lines, here set at 60 min) is shown in sub-figure (c). A clear global maximum at δt = 9 min is shown; this determines the time lag τij between the target (blue) and peripheral (red) time
series in sub-figure (a). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

1468
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Yij (δt ) = sgn(max(ρ (ki*,t ,ki*,t − δt ),δt )) [max (ρ (ki*,t ,ki*,t − δt ),δt )]2 (3)

with t ∈ TObs (t ,Δt ,Obs)

The indicator function contains the correlation raised to the power 2


(preserving the sign ( ± ) ), to make the distinction between good and
bad correlations more pronounced.
The time lag between two clear-sky index time series τij is de-
termined by the maximum of the indicator function Yij with respect to
the variable δt :
τij = arg max(Yij (δt ),δt ) (4)
This process of finding the time lag between two time series, is related
to cross-correlation, although in this case, not the entire time series are
compared: only a part of width Δt,Obs . A found time lag τij is char-
acterized by a quality factor, Qij , reflecting the goodness of the match
between the ki∗ (t ) and k ∗j (t −δt ) time series. Qij is defined by the max-
imum value of Yij (δt = τij ) , see Eq. (5).
Qij = Yij (δt = τij ) (5)
Finding time lags and associated quality factors of different peripheral
PV-systems with respect to the target PV-system is important for the
selection of the proper peripheral time series that can be combined into
a forecast of clear-sky index for the target PV-system. This selection Fig. 4. Example of selection of peripheral PV-systems. The target system is indicated by
process and the following forecast method will be discussed in the next the blue cross. The dashed line shows the found direction from where the clouds are
section and is visualized in Fig. 3(c). coming. The highlighted locations (gray circles) are the PV-systems that are found within
± 15 degrees from the cloud direction. The red circles indicate the locations that have
sufficiently high Qij and are included in the generation of the forecast. (For interpretation
3.1.2. Time lag optimizer of the references to color in this figure legend, the reader is referred to the web version of
Using Eq. (4), the found time lag τij can depend strongly on the this article.)
width of the observation window (Δt,Obs ), because of the inclusion of
different clouds. For this reason, instead of selecting a single time lag τij
At a selected moment t0 with associated Observation Window
as the result of a fixed TObs , the observation window length is varied by
TObs (t0,Δt ,Obs) , the selection algorithm determines the time lag τij and
scanning over δt for increasing values of Δt(temp)
,Obs ∈ {1,…,Δt ,Obs} . Each
corresponding Qij of the target system i with respect to the peripheral
TObs (t ,Δt(temp)
,Obs ) generates a pair of {τij,Qij } : from these the τij which scores PV-systems j that lie within the region determined by the found wind
the most cumulative quality “points” is selected as the one to be used,
direction ± 15 degrees (see Fig. 4). Additionally, a selection is made
along with the highest associated Qij . These points are calculated as the
based on cloud motion speeds. Although the actual cloud motion speed
sum of the quality factors belonging to each of the found options for τij
is not included in this research, some boundary conditions are set to
that are gathered in τx ∈ T as a weighted sum, see Eq. (6):
exclude non-physical speeds. When the distance between a peripheral
Δt ,Obs
PV-system and the target PV-system, projected onto the cloud motion
Score(τx ) = ∑ (Qij (Δt(temp) (temp)
,Obs ) | τij (Δt ,Obs ) = τx ), direction, and its assumed time lag τij lead to a cloud motion speed
Δt(temp)
,Obs = 1 (6) of > 150 km/h, the peripheral location is rejected. A lower boundary of
2 km/h is applied as well. Finally, a maximum distance of 15 km is
such that the final τij = (τx | Score(τx ) = max(Score(τx ),τx ∈ T )) . This
allowed for a peripheral location to be included in the forecast. An
way, a mixture of repetitive occurrence of found time lag and high
overall lower boundary condition of Qij ⩾ 0.4 was chosen. This value
correlation is rewarded with the emphasis on correlation, due to the
was chosen to include sufficient peripheral PV-systems for generating
power of 2 in Eq. (3).
the P2P forecasts, without including (too many) false or weak corre-
lations.
3.1.3. Cloud motion direction
This is an implementation of the assumption that a PV-system is
The entire Cloud Motion Vector, CMV (direction and speed) is not
influenced by clouds that come from one direction and it removes
explicitly calculated in the selection algorithm, but the direction can be
peripheral PV-systems that are too far away (i.e. the clouds changed too
found from the deviations of the assumption of perfectly correlated
much in shape, direction or speed) or peripheral PV-systems that are
lagged time series: the direction in which the kt∗,ij time series retain the
affected by local, incidental shading.
maximum correlation over the longest distance is the direction in which
the clouds are assumed to be advected [69]. The speed is implicitly
present in the found time lags between PV-systems at their respective
3.2. P2P forecast generation
distance. An EKO instruments (SRF-02) zenith-directed fish-eye total
sky-imager (TSI) (or cloud cam), located at UPOT1 was used to verify ∗
At the designated point in the future, t0 + FH, the forecast ki,̂ t0+ FH is
this method and comparison of the presented results in Table 6, see
constructed from all the available peripheral time series that have been
Appendix A.1. The cloud speed would require knowledge of the cloud
selected up to the moment t0 , under the condition that the time lag is
base height and that is not within the scope of this research: the P2P
sufficiently long, corrected for the time t at which the selection was
forecast method is aimed at using only measurements from the PV-
made. Under ideal circumstances (constant clouds and constant cloud
sensor network itself, supplemented by coarse temperature measure-
motion), the same set of peripheral PV-system j ∈ J (t ) is selected at
ments as the only exogenous data input. Further details of determining
each t. However, due to the changing nature of clouds and their motion,
the cloud direction procedure are described in Appendix A.2.
the set J (t ) can contain different peripheral locations j at different
moments t, of which the time lags can differ: to account for this mul-
1
Utrecht Photovoltaic Outdoor Testing Facility (UPOT), [86,87]. tiplicity, the j’s are labelled jt . These variations are incorporated in the

1469
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

used to validate the forecasts. All the error metrics are calculated using
the Δt = 60 s time series over the time series support period P that is
constructed of joining the time series of the chosen days into one time
series for a given choice of PV-system and FH.
Mean Bias Error (MBE):

̂ Ii ) = 1
MBE = mean(Ii − ∑ (Ii,̂t −Ii,t )
Nt ∈ P t∈P (10)

Relative Mean Bias Error (rMBE):


MBE
rMBE =
mean(Ii ) (11)

Root Mean Square Error (RMSE):

̂ Ii ) =
RMSE = RMS(Ii − ̂ Ii )2
mean(Ii − (12)

Fig. 5. Illustration of the single value of the Final Forecast ki,(̂ t0+ FH) as the black dot
indicated by the arrow. The shifted k ∗jt with jt ∈ J (t0) that are used in this example are Relative Root Mean Square Error (rRMSE)
shown as pink, thicker and dashed line segments; whereas the historic ones are shown as
̂ Ii )
RMS(Ii −
red, thinner and dotted lines. The blue line represents the target ki∗,t time series for re- rRMSE =
ference. (For interpretation of the references to color in this figure legend, the reader is RMS(Ii ) (13)
referred to the web version of this article.)
Forecast Skill (FS); where subscript (p) indicates clear-sky index
persistence:
forecast by taking the average value of all the available k ∗jt ,t0− τij . The set
J (t0) contains all previously selected jt , corrected for by time offset ̂ Ii )
RMS(Ii −
FS = 1−
t0−t that increases for times t in the past. If for example the correlation RMS(Ii (̂ p)−Ii ) (14)
between a particular j and the target PV-system i is sufficiently high for
subsequent t, this j will appear multiple times as jt , and will contribute Because the metric Forecast Skill (FS) ranges from < 0 (= bad forecast)
proportionally, as long as the thus shifted time series reaches up to (or via 0 (= equal to persistence forecast) to 1 (= perfect forecast), taking
beyond) the forecast horizon FH, see Eq. (7). arithmetic mean values of this metric is strongly biased towards nega-
tive values. Therefore, for aggregate results of multiple days or loca-
J (t0) = {jt ∈ J (t ) | t ⩽ t0 AND τij ⩾ (FH + t0−t )} (7) tions, we adopt the method of generating the Average Forecast Skill,
During periods of clear or overcast weather, the situation can occur (avgFS) via the method that is recommended in the same paper, and
that no peripheral PV-system is selected and available for forecasting. also adopted by Lorenzo et al. [70]: by taking avgFS as 1−s , with s the
In that case as well as the case for a chosen FH that is larger than the slope of the best linear proportional fit through scattered values of Root
largest available corrected τijt , there are no jt ∈ J (t0) and the P2P Mean Squared Error of (persistence-) forecast with respect to mea-
method chooses for clear-sky index persistence. surement: {RMSE(p) (π ),RSME(π )} , taken over equal-sized subsets
This process is described in see Eq. (8) and is depicted in Fig. 5. π ∈ P , see Eq. (15):
Average Forecast Skill (avgFS); where subscript (p) indicates clear-
∗ ⎧ jmean k ∗jt ,t0− τij when J (t0) ≠ ∅ sky index persistence:
P2P Forecast: ki,̂ t0+ FH = t ∈ J (t0)
⎨ k ∗ (t ) when J (t0) = ∅
⎩ i 0 (8) avgFSπ = 1−slope{RMSE(p) (π ),RSME(π )} (15)

The forecast time series ki,̂ t
will have the same temporal resolution Naturally, the value of avgFS converges to the definition of FS in Eq.
as that of the measured time series and spans from sunrise + 1 h + two (14) when the entire time series is included in one subset and the slope
times Δt,Obs until sunset − 1 h − one times Δt,Obs . is exactly equal to the ratio of the two values RMSE(P )/RMSE(p) (P ) .
Instead of taking subsets in the time domain, the average FS of different
3.3. GHI forecasts PV target systems can be calculated by adding tuples of
{RMSEx ,(p) (P ),RMSEx (P )} of the different PV target systems x.
Forecasting of clear-sky index kt∗ is a means to the final goal of Equivalently, x can indicate results of different forecast horizons if that
forecasting Global Horizontal Irradiance (GHI), IG,i,t . For the con- is the parameter of interest.
venience, the subscript “G” will be omitted in the notation for GHI:

IG → I . GHI is obtained by multiplying the found ki,̂ t by the modeled 3.5. Daily irradiance characterization
clear sky irradiance Ic,i,t :

Ii,̂t = ki,̂ t ·Ic,i,t (9)
The irradiance on a day d can be characterized by the total irra-
diance compared to the clear sky irradiance, or Daily clear-sky index, Cd
This way the reported error metrics can easily be compared to other see Eq. (16) and the irradiance variability compared to clear-sky
studies of solar (irradiance) forecasting. variability, or Relative Variability Vd see Eq. (17). This gives quantitative
measures to differentiate broken cloud days with e.g. clear sky days or
3.4. Error metrics overcast days, see Fig. 9. Cd and Vd are determined using measurement
data from calibrated pyranometers at the UPOT. UPOT is located in the
Verification of the P2P forecasting method was done comparing the center of the PV-Sensor Field and because the variability is determined
forecast Ii,̂t for a target PV-system with the actual recorded Ii,t at that over the entire day, the value can be taken as representative for the
PV-system (i) . Relative Root Mean Square Error (rRMSE) of the fore- daily variability of any of the PV-systems. This is under the assumption
casts and comparison of the forecast to using only clear-sky index that the general weather is homogeneous throughout the PV-Sensor
persistence can be done, to provide the so called forecast skill [88]. The Field which roughly covers the Province of Utrecht (NL) and measures
following equations give an overview of common error metrics that are approx. 1400 km2 .

1470
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 6. Left. Example of theP2P forecast (black line) on a variable day, compared to the measured GHI time series (blue, dashed line). Right. Graph showing the rRMSE of the forecast
(black, dashed line) and of the persistence forecast (orange, dash-dotted line) for different values of FH, along with the resulting forecast skill (red, solid line). The rMBE is indicated by
the gray, dotted line. The vertical line shows the 13-min mark of which the time series is shown in the graph on the left. (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)

∑t ∈ d It the target systems is fixed and does not change per day or due to the
Cd =
∑t ∈ d Ic,t (16) variable CMV direction in order to attribute variations in the perfor-
mance of the forecasting method to the irradiance characteristics and
∑t ∈ d (ΔIt )2 not to variations in the number and location of test systems. The
Vd = measurement data was taken at an interpolation interval of Δt = 60 s .
∑t ∈ d (ΔIc,t )2 (17) The maximum length of the Observation Window was chosen to be
Note that this definition of relative variability is subtly different Δt,Obs = 60 min , so the maximum FH is in principle 60 min as well.
than that of the Variability Index by Stein et al. [89]: the addition of However, it only rarely occurs that a high quality time lag of this length
+ Δt to the ΔI in the latter is mathematically ambiguous as it is a line is found, so the FH’s studied in this paper range from 1 to 30 min
length constructed from quantities that have different units (time and Specifically, the best FS is usually found in the range between 5 and
irradiance fluctuation). Nevertheless, by construction either description 8 min, so this range will be highlighted as the most interesting one for
classifies variability comparably. Mazorra Aguiar et al. [90] use the the P2P method.
standard deviation of clear-sky index fluctuation, which is another,
comparable way to assess daily irradiance variability. This reduces to 4.2. P2P forecasting results
RMS as described in Eq. (17), because the average value of the fluc-
tuations is necessarily 0 as the value of kt∗ starts at 0 and returns to 0 The P2P algorithm produces a forecast time series at the same
during an entire day. temporal resolution as the input data. Fig. 6 shows an example of a P2P
forecast with generally positive FS on a day with broken clouds
4. Results (Cd = 0.68,Vd = 112 ), for which results are presented in Table 2. Fig. 7,
and Table 3 show an example of a P2P forecast on a day with moderate
The P2P-forecasting algorithm was run over a series of selected days variability (Cd = 0.82,Vd = 32 ). As can be seen from the graph in Fig. 7
with high irradiance variability (due to intermittent clouds), as well as (left), especially in the morning, the clouds were easy to follow; later on
over a longer period of time that also includes a mix of (partly) overcast the day the forecast remains reasonably well on its own, but with
and clear sky days. The results are compared to other reported fore- rRMSE comparable to that of the clear-sky persistence and thus with
casting algorithms that have similar data resolution and forecast hor- slightly negative FS. Fig. 8, and Table 4 show an example of a P2P
izon. forecast on a mostly clear day (Cd = 0.96,Vd = 4 ). This forecast has a
negative FS, but as is visible from the graphs, the actual forecast time
4.1. Boundary conditions series follows the measured time series quite well. The definition of the
FS punishes this increase of (r)RMSE as the (r)RMSE of the clear-sky
The algorithm was run over the six month period of March–August index persistence forecast will be very small over the day.
2015 to assess the average performance over different weather types. In general, FS as a function of FH tends to start out low (negative).
This range includes 182 days, as the 13th and 14th of July were un- This is due to the fact that at very short FH, clear-sky index persistence
available due to a disconnected server. The 11 chosen target PV-systems becomes harder to beat. With FH larger than a few minutes, however,
are located near the center of the PV-Sensor Field to ensure ample the P2P method generally shows improvement over persistence by
peripheral PV-systems from all wind directions, see Fig. 1. The choice for combining time lagged peripheral k ∗j time series. Finally at high FH,

Table 2
Rounded error metrics for an example PV-system on the 30th of March 2015 (Vd = 118) at chosen values of forecast horizon (FH).

FH [min] 1 5 9 13 17 21 25 29
MBE [W m−2 ] 0.00 −0.01 −0.01 −0.02 −0.05 −0.08 −0.07 −0.06
rMBE [%] 0.40 −1.10 −1.43 −3.56 −6.73 −11.4 −10.4 −7.78
RMSE [W m−2 ] 0.26 0.28 0.29 0.31 0.33 0.35 0.36 0.35
rRMSE [%] 38.9 41.3 41.7 45.0 46.3 49.5 50.6 48.9
FS [%] −35.5 9.45 16.2 16.3 15.8 4.31 2.28 3.95

1471
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 7. Same set-up as Fig. 6 for a moderately variable day and a different location. The vertical line shows the 9-min mark of which the time series is shown in the graph on the left.

close to the maximum, there are less and less available peripheral time this paper as when Vd ⩽ 75 or Vd > 75, based on the cross-over of cu-
series k ∗j,t with j ∈ J (t0) (see Eq. (8)) that span up to this FH and the mulative (binned) forecast skill from FS < 0 into FS ⩾ 0 in Fig. 10(a),
P2P method automatically reduces to persistence, resulting in FS = 0 and is used to show the dependence of FS versus Vd in Table 6. In the
per definition at the maximum FH = Δt,Obs . The specifics of the FS total measurement period, 126 days were classified as low Vd days and
versus FH behavior is different for each day as the forecast time series 56 days were classified as high Vd days. For Cd , a relation to FS is not as

kt ̂ depends on the set J (t0) at each evaluated moment, which may be clear, as the average FS is fairly constant except for values near the
different for each FH. This is visible in e.g. Fig. 8 (right). minimum and near the maximum, see also Fig. 10(b): the FS tends to be
To investigate the statistics of the behavior of the P2P method, the lower for either very clear or very overcast days. This is a reflection of
results covering the 11 centrally located PV-systems are shown in the moderate interdependence of Vd and Cd , which is also visible in the
Tables 5 and 6. These tables show that on average, the P2P-method has shape of the {Vd,Cd} scatter plot in Fig. 9. Basically, the charts of Fig. 10
a higher FS on days with a high relative variability Vd , compared to all can be interpreted as side-views of the non-averaged data in the plot
days, or to days with low variability. The MBE, is on average, modest points in Fig. 9.
around 3%. Especially the RMSE is rather high (in absolute terms). This As discussed before and visible in e.g. Fig. 11, there is a tendency for
is the result of the fact that we measure at very high temporal resolution the FS to be very low at low forecast horizons FH< 5 min and for FH
and missing some details of a small shade or over-irradiance event values close to the maximum of FH = 30 min The rRMSE is found to
passing, is punished very strongly. Relative to the clear-sky index per- increase from 25% to 35% for the Low Vd days and from to 40% and
sistence forecasts, these RMSE values remain comparable (or even 60% for the High Vd days over the studied forecast horizons, see Fig. 12.
better in the case of the highly variable days) which is reflected in the The results that combine all the studied days are in between these. From
FS. The distribution of total FS values per PV-system, per FH is gra- the positive avgFS results, it is known that the rRSME of the persistence
phically presented in Fig. 11. forecast is higher than that of the P2P forecast, especially for the High
Vd days: this is the desired improvement.
4.3. Dependence of results on daily irradiance variability
5. Discussion
Days that are characterized by broken, intermittent clouds naturally
show high irradiance variability. The P2P forecasting algorithm per- The P2P forecasting method was evaluated over a subset of days
forms better on variable days than on clear or overcast days, because of with either high, all or low irradiance variability. The average forecast
the need to relate shadow transients from moving clouds between dif- skill (avgFS) was over these periods respectively, indicating the su-
ferent PV-systems. This dependence is shown by differentiating the perior performance of the P2P method over persistence during the
results between all days in the measurement period, and a selection of highly variable days, which are the most interesting from the per-
days that show high variability, see Fig. 9: a so-called Arrowhead Dia- spective of an electricity grid.
gram that owes its name to the triangular shape of the scatter plot when A downside of P2P forecasting is that forecasts can only be made for
it is populated by days covering a diverse range of weather conditions the PV-systems in the interior of the area spanned by the PV-systems, as
[89]. Vd and Cd cover a large range of the diagram, validating the choice upwind PV-systems/sensors are necessary as input for the forecast.
for the measurement period. Note that the highest of Cd occur at days Forecasts for locations in between the PV-systems can in principle be
with few clouds, as the effect of extra diffuse (super) irradiance from a generated, using spatial averages of the PV-systems, this will be in-
small amount of clouds contributes positively overall to the total irra- vestigated in further research.
diance. The cross-correlation of the dispersed PV-systems was used on its
Distinction between low and high irradiance variability is defined in own to determine the cloud direction as attempted by e.g. [91,78], i.e.

Table 3
Rounded error metrics for an example PV-system on the 11th of May 2015 (Vd = 32) at chosen values of forecast horizon (FH).

FH [min] 1 5 9 13 17 21 25 29
MBE [W m−2 ] −6.86 −8.52 −8.25 −9.25 −9.69 −11.2 −13.4 −11.7
rMBE [%] −1.24 −1.54 −1.49 −1.68 −1.75 −2.03 −2.42 −2.13
RMSE [W m−2 ] 50.4 69.4 75.8 118 117 116 120 115
rRMSE [%] 9.13 12.6 13.7 21.4 21.1 21.1 21.7 20.8
FS [%] −15.7 21.4 29.3 −1.77 −1.86 −1.82 −2.65 −0.58

1472
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 8. Same set-up as Fig. 6, on a clear day for the same location as in Fig. 6. The vertical line shows the 4-min mark of which the time series is shown in the graph on the left.

circumventing the input of the cloud cam. Due to the inhomogeneity of means that for clear or overcast days with low variability, persistence
the spatial distribution of PV-systems, assessment of cloud motion forecasting is preferred and this should then be the choice. The ex-
vector (CMV) direction through the PV-sensor network itself can be pected variability can, in principle be inferred from coarser medium-
biased in directions with higher density of PV-systems. The P2P method term (day-ahead) forecasting, e.g. through satellite or numerical
was also run, using the cloud cam determined CMV direction. The re- weather prediction forecast methods.
sults are presented in Table 6 and higher FS were found in any of the Furthermore, the measurement data was down-sampled from ≈ 2 s
cases. The downside to using a cloud camera is that it has to be avail- to mean bins over 60 s intervals. These choices were based on experi-
able and processed, whereas the P2P method as presented here can rely ence and intermediate results during the development of the forecasting
purely on measurements of the PV-systems. The benefits of using a algorithm as well as on optimizing the calculation time for generating
cloud cam over the (sensor) network-based de-correlation method are: the results.
high, isotropic directional resolution that is not limited by the isotropy
of the sensor network; the ability to determine the CMV during overcast 5.1. Comparison to other forecast studies
periods; the possibility to differentiate between different cloud layers
(depending on the presence of a cloud-classification algorithm, and In general, reported error metrics depend strongly on the chosen
feasible if cloud base height data from an associated ceilometer is methods and data specifics (temporal resolution of input data, forecast
present). Table 6 shows that the cloud cam determined CMV direction horizon, forecast resolution, measurement period). Usually only best-
resulted in better forecast skill (FS) in the case of the P2P forecast al- practices or mean values are reported which makes it difficult to di-
gorithm. Contrary, the benefits of the network-based method are: no rectly compare results from different studies. The methods that are
extra device and image processing necessary; relatively robust for shown in Table 7, are based on studies mentioned in Section 1.3 and are
minor sensor outages; guaranteed cover of the entire sensor network; the ones with the most similar set-up and boundary conditions as the
and correspondence of method for CMV and forecast generation. During P2P method. Broader comparison of FS versus FH results can also be
extreme clear sky conditions, or otherwise difficult circumstances that found in an extensive review on solar forecasting methods by Anto-
do not generate clear CMV using either option, no CMV might be de- nanzas et al. [21].
termined. In that case, forecasting is usually better done by (clear-sky It should be stressed that in Lipperheide et al. [77], the cloud speed
index) persistence anyway so lack of CMV knowledge is irrelevant. persistence method is assessed and the 5 sensors that generate the cloud
A regular grid of identical GHI sensors rather than PV-systems could motion vector are mentioned as the “sensor network”. The spatial dis-
be used as input for the P2P forecasting algorithm that is described here tribution of the sensors of each of the shown studies differs based on
and circumvent errors that arise from the use of an inverted decom- geography and available resources, nevertheless the sensor density
position/transposition model to convert AC Power into GHI. Also, ac- gives a good context to the results, next to the used temporal resolution
tual GHI sensors would eliminate the need for quasi real-time ambient Δt . Aryaputera et al. [74] provide a method to forecast for unobserved
temperature information as an input for the conversion of power into locations by systematically apply the found (stationary) spatio-tem-
GHI. Although (measurement and) conversion errors are present in the poral correlations in the sensor network. Both the results for observed
current data collection, the principle and the general trend of the results and unobserved locations are shown in Table 7. Furthermore, each of
of the P2P forecasting algorithm itself are the interest of our present the reported method has a different view on the construction and ex-
study. ecution of the forecast horizon (FH) and calculation of aggregated FS
The results show that the P2P forecasting method works best on results. Relating the former, Yang et al. [39] and Aryaputera et al. [74]
days characterized by high daily irradiance variability. The fact that choose to first resample the time series to the desired FH and then
forecasting performance decreases with variability is not an issue; this forecast one step ahead. This is quite different to Lonij et al. [69] and

Table 4
Rounded error metrics for an example PV-system (same one as in Table 2) on the 30th of June 2015 (Vd = 4) at chosen values of forecast horizon (FH).

FH [min] 1 5 9 13 17 21 25 29
MBE [W m−2 ] −0.06 −0.05 −0.04 −0.03 −0.02 −0.009 −0.0005 0.008
rMBE [%] −6.67 −5.79 −4.74 −3.45 −2.23 −1.06 −0.05 0.98
RMSE [W m−2 ] 0.12 0.12 0.11 0.11 0.11 0.10 0.11 0.11
rRMSE [%] 13.7 13.4 13.0 12.7 12.3 12.1 12.4 13.1
FS [%] −647 −150 −79.3 −47.9 −31.0 −21.7 −17.4 −15.5

1473
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Table 5
Table showing the (r)MBE and (r)RMSE averaged over the 11 central PV-systems, specified over the range of included FH on the High Vd days. The value between brackets shows the
sample standard deviation.

FH range [min] {1 …4} {5 …8} {9 …12} {13 …16} {17 …20} {21 …24} {25 …28}

MBE [W m−2] High Vd 15.9 (15.3) 17.0 (14.5) 15.1 (13.4) 12.0 (11.9) 8.82 (10.9) 4.78 (9.41) 2.15 (7.22)
All 10.4 (17.3) 10.6 (16.4) 9.45 (15.2) 8.01 (13.8) 6.24 (12.4) 4.25 (10.8) 2.87 (8.78)
Low Vd 7.82 (18.6) 7.50 (17.7) 6.74 (16.4) 6.05 (14.9) 4.98 (13.4) 3.95 (11.7) 3.12 (9.84)

rMBE [%] High Vd 3.76 (3.60) 4.02 (3.44) 3.57 (3.19) 2.86 (2.85) 2.10 (2.61) 1.14 (2.27) 0.53 (1.76)
All 2.65 (4.36) 2.68 (4.14) 2.40 (3.85) 2.04 (3.49) 1.59 (3.17) 1.09 (2.76) 0.74 (2.26)
Low Vd 2.05 (4.86) 1.97 (4.62) 1.77 (4.30) 1.60 (3.91) 1.32 (3.54) 1.05 (3.09) 0.83 (2.61)

RMSE [W m−2] High Vd 189 (24.6) 213 (25.2) 227 (25.3) 237 (26.0) 244 (26.9) 250 (28.1) 254 (28.9)
All 137 (18.5) 152 (18.3) 162 (18.2) 169 (18.3) 174 (18.4) 178 (18.8) 181 (19.2)
Low Vd 106 (17.2) 115 (16.0) 122 (15.2) 127 (14.6) 131 (14.1) 134 (14.0) 136 (14.1)

rRMSE [%] High Vd 44.8 (4.53) 50.3 (4.34) 53.7 (4.22) 56.0 (4.26) 57.8 (4.36) 59.4 (4.46) 60.4 (4.61)
All 35.2 (4.37) 39.0 (4.12) 41.4 (3.95) 43.3 (3.84) 44.7 (3.80) 45.7 (3.81) 46.6 (3.89)
Low Vd 28.1 (4.55) 30.5 (4.16) 32.3 (3.85) 33.8 (3.59) 34.9 (3.40) 35.6 (3.36) 36.2 (3.40)

Lorenzo et al. [70] who translate the cloud motion as desired, as far
ahead as possible. Although similar in set-up, Lipperheide et al. [77] is
limited by the fact that the sensors are within the measured inverters
and the spatial extent is ≈ 1 km2. With our current P2P method, we fix
the forecast horizon by choosing from available intermediate forecasts
as described in Eq. (8). The final forecast is produced via the mean of
the available intermediate forecasts. This means that any in-
homogeneities in the Inverse PV model’s operation from P to GHI (to k ∗)
are smoothened and that the absolute offset of a few peripheral k ∗j,t − τij
values (through unfortunate local shading or reflections) do not trans-
late as strongly as they might when only the best correlating sensor is
chosen (as is the case in Yang et al. [39] and Achleitner et al. [76]).
Assuming correct time lag selection, this averaging does not equate
directly to temporal smoothing, rather some combination of lagged,
upwind spatial smoothing. This is a different manner of smoothing than
Lorenzo et al. [70], who first interpolate the irradiance field before
moving it in the direction of the cloud motion vector (apart from
comparing advected forecasts to spatial average persistence). However,
these considerations are also under the assumption that clouds retain Fig. 9. Characterization of the days in the measurement period (March–August 2015).
Each day (dot) is colored by the average forecast skill (avgFS) ranging over the locations
their shape and velocity.
and over FH ∈ {5,…,8} . The color scale ranges from −50% to +50%. Yellow into green
The P2P method does not perform as well as we expected from the
into blue indicates more positive avgFS, whereas gray into red indicates more negative
fact that it literally uses the clouds that are upwind to forecast for the avgFS. Negative values beyond <−50% fade into black: in particular a few days with very
near future. This means that the assumption that the clouds retain their low Vd . The insets show example daily GHI(t) graphs that correspond to the descriptions.
shape and velocity is perhaps too strong for the geographical range (For interpretation of the references to color in this figure legend, the reader is referred to
studied (tens of kilometers). The relation between actual cloud shadow the web version of this article.)
shape, its deformation over time and the applicability of network-based
forecast methods like the P2P method will be the focus of further re- power generated by a network of PV residential systems. The P2P
search. forecasting method is suited to forecast short-term, highly variable ir-
radiance at the same temporal resolution as the input data or could be
6. Conclusion used to refine longer time-ahead coarse forecasts. The P2P method
shows an average forecast skill of 5.99%, −1.61% and −16.0%, over
We have developed a P2P forecasting method as a means to provide forecast horizons between 5 and 8 min for the days characterized by
short-term, local GHI forecasts to be used in forecasting variation of high, all and low relative variability. These results align to some extent

Table 6
Results of avgFS [%] for daily {RMSE(p) ,RMSE} -tuples ranging over all the 11 central PV-systems, specified over the range of included FH and included days. The values between brackets
indicate the R2 of the fit that was used to generate the slope for the avgFS. The top three rows show the results of the CMV determined through the de-correlation method, the bottom
three rows show the results obtained via the cloud cam.

FH range [min] {1 …4} {5 …8} {9 …12} {13 …16} {17 …20} {21 …24} {25 …28}

De-correlation CMV
High Vd days −4.17 (0.955) 5.99 (0.984) 5.41 (0.989) 3.81 (0.991) 1.96 (0.993) 0.58 (0.994) 0.45 (0.995)
All Vd days −15.7 (0.869) −1.61 (0.930) −0.50 (0.947) −1.05 (0.957) −1.99 (0.965) −2.58 (0.971) −2.47 (0.976)
Low Vd days −41.3 (0.819) −16.0 (0.890) −10.6 (0.911) −8.72 (0.926) −7.87 (0.937) −7.12 (0.947) −6.58 (0.957)

Cloud Cam CMV


High Vd days 0.15 (0.956) 8.95 (0.985) 7.72 (0.989) 5.39 (0.991) 3.08 (0.994) 1.40 (0.995) 1.13 (0.996)
All days −11.0 (0.884) 2.03 (0.945) 2.71 (0.962) 1.64 (0.973) 0.33 (0.980) −0.58 (0.985) −0.56 (0.989)
Low Vd days −38.5 (0.848) −12.6 (0.916) −6.85 (0.938) −5.05 (0.952) −4.29 (0.964) −3.82 (0.973) −3.29 (0.979)

1474
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. 10. (a) Plot of the FS over all the PV-systems and FH ∈ {5,…,8} ; the values are generated from joined time series, containing all days of which the Vd falls into a specific bin to show the
trend in dependence of FS versus Vd . The medians are indicated by the white bars. Outliers beyond 1.5 (and 3 times) the inter-quartile-range are shown as (small) dots. (b) Same data, but
now ordered according to binned Cd values. At the 0 ⩽ Cd < 0.1 bin, no values were present.

Fig. 11. Box-Whisker plots of the forecast skill (FS) of the 11 central PV-systems per
Fig. 12. Box-Whisker plots of the rRMSE of the 11 central PV-systems per forecast horizon
forecast horizon (FH). Lighter boxes show FS calculated over P = 126 Low Vd days, the
(FH), same layout as Fig. 11.
orange ones cover all the days and the darker ones cover only the P = 56 High Vd days.
The median at each FH value is indicated by the white bars in the boxes. Outliers beyond
1.5 (or 3) times the inter-quartile-range are shown as (small) dots. significantly higher forecast skill during days with high variability,
compared to days with low variability (either clear or overcast). This
with comparable other methods in the field, but are not as good as those means that, although the relative RMSE (rRMSE) might increase, the
of other researchers that incorporate more knowledge into their fore- rRMSE of the benchmark clear-sky index persistence increases more
casts, e.g. known cloud motion vector or trained correlation data. strongly, on average and especially for forecast horizons in the range
However, the method we present provides a case study of a generic 5–8 min When it is known that a clear or overcast day can be expected
method of forecasting based on cross-correlation that only uses locally (e.g. from longer-term weather forecasting), simple clear-sky index
available data, not based on historic training or inference with the persistence forecasting suffices. Knowledge of the detailed PV power
benefit of a fixed forecast horizon that is independent of the temporal output and its fluctuations is relevant for the integration of more PV
resolution of the data. into the (future) smart electricity grid.
Furthermore, the numerical results show that the P2P method has

1475
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Table 7
Comparison of the P2P forecasting algorithm to other deterministic, network based short-term forecasting methods on variable/cloudy/non-clear days. The values are maximal values
and/or mean values reported in or calculated from data in the given references. The references included are: P2P = this work, I = [69], II = [70], III = [71], IV = [74], V = [77], VI =
[76]. The method used in the reference is indicated by “CMV” for known cloud motion vector from weather data, and “SI” for any kind of sky imager/cloud cam. Data type refers to the
processed data and (if applicable) a conversion to compute error metrics.

Ref. Data type Meas. period Δt Sensor density FH FS

P2P k∗ into GHI 56 days 60 s 133


≈ 0.095 km−2 5–8 min average FS 5.99%
1400 km2
P2P k ∗ into GHI + SI 56 days 60 s 133
≈ 0.095 km−2 5–8 min average FS 8.95%
1400 km2

I P into k ∗ + CMV < 1 year 15 min 80


≈ 0.032 km−2 15–90 min −8.1–5.3%
2500 km2
II k∗ + CMV 46 days 1s 12 + 3
≈ 0.011 km−2 1–120 min average FS 19–23%
1400 km2
III k ∗ into GHI + CMV 13 days 10–300 s 17
≈ 14 km−2 var (see Δt ) var. 40–0%
1.2 km2
IV k∗ into GHI 13 days 10–300 s 17
≈ 14 km−2 (var.), reported: 50 s max. unobs. 36%; obs. 43%
1.2 km2
V P 171 days 1s 5 var. {20; 60} max 180 s {16.2; 10.6}%
≈ 5.6 km−2
0.9 km2
VI P + SI 5 days 5s 19
≈ 19 km−2 var. < 2 min max 97%; mean 93%
1.0 km2

Acknowledgments partly financially supported by the Netherlands Enterprise Agency


(RVO), through funding the project Solar Forecasting & Smart Grids
The authors would like to thank Paul Raats, Bas Vet, Santiago (SF & SG) within the framework of the Dutch Topsector Energy, “TKI
Peñate Vera (DNV-GL), Lou Ramaekers (Ecofys) and Maurits Uffing (UU Switch2SmartGrids” under TKISG02017. The authors would especially
Geo-ICT) for fruitful discussions and support. Furthermore, thanks to like to thank the participants in the SF & SG project who volunteered to
Atse Louwen and Arjen de Waal (UU) for the UPOT data. This study is take part in the PV-Sensor Field and had their PV data measured.

Appendix A. Cloud motion vector direction determination

The background to the two methods of determining the Cloud Motion Vector (CMV) direction is treated here.

Fig. A.13. Geometrical correspondence of an image pixel in the fish-eye


photograph (black star) at fractional radius ri , to the actual situation at
cloud base height h at radius ri′ (blue star). The maximum FOV is illustrated
by the red dot-dashed line. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

Fig. A.14. Three snapshots of the flattened cloud


cam image on 9 July 2015 and the resulting CMV
direction of movent (yellow). These (sky) images
are reflected in the horizontal axis such that the
azimuth is as expected on a map with the North at
the top. These images correspond to the time
stamps a, b and c in Fig. A.17. (For interpretation
of the references to color in this figure legend, the
reader is referred to the web version of this ar-
ticle.)

1476
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. A.15. Example of pure temporal de-correlation for one fixed location on 9 July 2015, halfway during the day. Blue (Δ) kt∗ lines for the target location i, red lines for the time lag
shifted peripheral locations j. Left column shows time lag τij and maximum correlation Cij four upwind peripheral locations that at increasing distance. The right column shows the same
location pairs, now for Δkt∗ based on maximum ΔCij . In all cases, the τij is determined based on the observation window TObs = 1 h between the vertical lines. Note that the time series
retain comparable shape, even outside the observation window indicating constant cloud speed. Deviations are the result of changing cloud shape, number and also possible non-perfect
PV meta data and Inverse PV model effects. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

A.1. Cloud camera

At UPOT, a Total Sky Imager (TSI) or cloud cam is present. The cloud cam that is used, is an EKO instruments (SRF-02) zenith-directed fish-eye
camera that takes two successive (1600 × 1200) pixels JPEG (RGB) photographs every 10 min These photo duo’s consist two photographs of shutter
speed 1/2000 s and 1/500 s; and are taken 14 s apart. Photographs of the sky are distorted by the lens in order to project the sky dome hemisphere,
centered at the Zenith to a small disk-shaped region on the camera’s Charge Coupled Device (CCD). Translation from points in the sky (at some fixed
height in the case of the bottom layer of clouds) to points on the image is described by the here called fish-eye projection function.
The original fish-eye function (Eq. (A.1)), which is necessarily a bijection over the input range that maps object pixels uniquely to image pixels,
can be thought of as two parts: the mapping from the cloud layer (object or plane) pixels at radius ri′ to the view angle θi , and the mapping from the
view angle to the photograph (image plane) pixels ri (as a fraction of the maximum camera CCD radius R), illustrated in Fig. A.13. This construction is
rotationally symmetric around the vertical axis, in other words, independent of the polar coordinate φ in the photograph.
Φ(r ′) = (ϕlens ∘ϕprojection )(r ′): r ′ → θ → r (A.1)
The projection function assumes a flat cloud layer that remains at the same height at least over the range of interest (up to the maximum Field-Of-
View (FOV) angle): it is thus from basic geometry that this function is described by Eq. (A.2).
r′
ϕprojection (r ′) = arctan ⎛ ⎞
⎝h⎠ (A.2)
The second step is the mapping of view angles θ to photograph radii ri and it involves the (unknown) and non-ideal lens function. Instead of

1477
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. A.16. Temporal de-correlation examples on a day with variable irradiance (9 July 2017), in three direction bins (left to right: −55°,−145° and 125° ) and at three time stamps (top to
bottom: 8:02, 10:32 and 13:02 UTC). De-correlation of the distance bin average maximum Δk ∗ correlation ΔCij is fitted to the PV-system separation distance dij under the assumption of
unknown but constant CMV. The left graph shows the direction in which the correlation derates the slowest: this is in the direction of the de facto found CMV origin. The middle graphs
show the cross-wind direction. In the right graphs, the ΔCij derates completely within a few distance bin steps: this happens to be the down-wind direction bin.

looking for this lens function, and having to invert it later on, we opted to search for the inverse lens function directly as it is the total inverted fish-
eye transformation function Φ−1 (r ) , that maps photograph pixels to cloud plane pixels, that we are interested in. These steps are displayed in Fig.
A.13.
Φ−1 (r ) = (ϕlens ∘ϕprojection )−1 (r )
−1 −1
Φ−1 (r ) = (ϕprojection ∘ϕlens )(r ) (A.3)
−1 −1
A few standard candidates for the function ϕlens (r ) were tried: equidistance projection ϕlens (r ) = (π /2)(r ) and orthogonal projection
−1
ϕlens (r ) = arcsin(r ) , but these did not produce sufficiently flat image transformations. The following odd polynomial function, see Eq. (A.4) showed
to be a very good inverse lens transformation for our particular fish-eye camera. For discussion and more general methods concerning fish-eye lenses,
see e.g. Kannala and Brandt [92] or Urquhart et al. [93].
−1 π
ϕlens (r ) = (4r + r 3 + 2r 5) with r ∈ [0,1]
14 (A.4)
Because this is a forward image transformation, the missing output pixels in the flattened image are linearly interpolated. A maximum FOV angle
with respect to the vertical of 70° is set, which is translated to a fractional radius of rmax = ϕlens (70°) ≈ 0.895 of used pixels in the input photo. This
value was numerically determined, as the inverse of the inverse lens function, (ϕlens) , does not exist because ϕlens −1
(r ) (Eq. (A.4)) is a fifth-order
polynomial.
In the forward image transformation, Eq. (A.5), this limit is implemented by curtailing r. The polar angle is mapped to itself: φi′ = φi . When
applied to the discrete sets of pixels, the coordinates {φi′,ri′} are rounded to the nearest discrete pixel.

⎧ Φ−1: {φi,ri} → {φi′,ri′}


−1
⎨ {φi′,ri′} = {φi,tan(ϕlens (min(ri,0. 895…)))} (A.5)

The flattened image is rescaled to a radius of 200 pixels to retain sufficient details but making calculations not too comprehensive. Position and
motion in the flattened image is now directly related to actual motion at the height of the clouds, more specifically in the view-direction of the
clouds. Over the studied spatial range (max. 60 km), the studied region at constant cloud base height can be considered flat. For the calculation of the
direction of clouds, absolute scale is not necessary, so the cloud height h is set to h = 1 and r ′ is interpreted as a fraction of the size of the flattened
image. For the purpose of pre-selection of upwind peripheral locations in the P2P algorithm this is sufficient.

A.1.1. Cloud motion detection


The photos from the photo duo’s with different shutter speed are ideal for deriving the cloud motion, because the short interval of 14 s is enough

1478
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Fig. A.17. The top row (a–c) shows three different chart pairs that plot instantaneous bC [km] (top) and R2 of the fit as the radius of the sector charts differentiated to each 10° direction
bin on 9 July 2015 (on 08:00, 10:30 and 13:00 UTC). The selected CMV direction origin is colored light yellow. The R2 charts are colored as gray below R2 = 0.75 and yellow into green
for 0.75 ⩽ R2 ⩽ 1. The bottom left graph shows the found CMV direction origin of the de-correlation method (full line) and the CMV direction origin of the cloud cam (dashed line), as in
Fig. A.14. The dashed vertical lines indicate the time of the charts a, b and c. The red, solid line denotes the time (at 16:20 local time = 14:20 UTC) that marks the end of the cloud cam
images and from which point on the cloud cam CMV direction origin values are extended as a straight line. The bottom right graph shows the same lines, on a magnified vertical axis. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

to show motion, without significant change in the shape and structure of the clouds. The two photos from a duo are aligned in terms of their color
histogram to make it easier for direct comparison of the features. Of the duo’s, the image of the brightest one (1/500 s) is taken as the reference.
The ImageFeatureTrack functionality of Wolfram Mathematica® [84], is used to track key points in two commensurate images. These key points
are locations in the image that have distinct spatial entropy (sharp edges, corners, color gradients, etc.). The resulting list of starting and end points
of the tracked features between two successive images gives, on average, a very good approximation of the (quasi-static) direction of the clouds.
Using this technique on the raw RGB photographs introduces some artifacts: the sun, along with lens flares and reflections, is a very prominent
feature that does not need to be tracked. Also, the inverse fish-eye transform may distort or exaggerate the inferred motion vectors, especially near
the edges of the FOV. Adding the brightest pixels (the sun and most of the lens flares) to the static mask. These pixels are then ignored by the tracking
algorithm. Some of the bright pixels on the clouds are removed, but this does not interfere with the determination of the overall motion, assumed
that the motion is approximately uniform and dominated by the most pronounced clouds. The motion vectors are determined in the flattened
photographs, see Fig. A.14 as an example. The size of the found direction vector (yellow) is for illustration purposes only; the absolute value of the
motion (pixels per 14 s) is stored for optional speed determination in combination with known cloud height. Comparison of the results to another
CMV detection method (through de-correlation) is treated in the next section.

1479
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

A.2. Temporal de-correlation

To what extent does the assumption of static and quasi-statically moving clouds that is the basis for the P2P forecasting algorithm hold? This is
investigated by comparing lagged clear-sky index measurements of real clouds, over limited time periods: see Figs. 3 and A.15. This is under the
assumption that the cloud motion speed does not change significantly within that period and hence the time lag is expressed as distance lag: it
describes how the clouds change in appearance while moving from PV-systems separated by dij = vc τ cosαij , with αij the angle between the system
separation vector and the Cloud Motion Vector (CMV).
Comparable to the Quality factor Qij as in Eq. (5), the value (Δ) Cij is defined as the maximum found correlation ρ of two PV-systems (i and j) time
series: either of the clear sky index kt∗ itself or of its fluctuations, defined by Δki∗,t ≡ ki∗,(t + dt ) −ki∗,t , scanning over lags δt within an observation window
TObs , see Eq. (A.6).
(Δ) C =
ij
max(ρ ((Δ) ki∗,t ,(Δ) k ∗j,(t − δt ) ) | t ∈ TObs (t ,Δt ,Obs),δt ) (A.6)
At time t, time lag τij and corresponding correlation factor (Δ) C
are calculated for all possible PV-system pairs i and j, following Eq. (4) re-
ij
spectively. The difference with the P2P method is that the CMV direction origin is determined through ΔCij , which is based on Δkt∗, whereas Qij is
based on kt∗. Per bin of 10° of separation direction, the values of ΔCij are plotted against the PV-system separation distance dij . The conclusion is that
the empirical correlation decreases when PV-systems are further apart: indication of the temporal derating of the assumption that the clouds remain
equal in shape and size (initially reflected in the assumption that the clear-sky index time series ki∗,t and k ∗j,t are equal, lagged copies). This derating of
correlation (de-correlation) is averaged over distance bins of 1 km centered halfway through rounding, to retain only the coarse trend, and then fitted
by an exponential model, to capture a characteristic length scale, see Eq. (A.7). The auto-correlation of ρ (d = 0) = 1 is not visible due to rounding in
the lowest bin, but is forced through the fit model construction.

Fig. A.18. Density plot of scattered 10 min CMV direction origin determined through the cloud cam and via the de-correlation method on the (Δ) kt∗ data in the measurement period
March through August 2015, corresponding to the measurement period of the P2P forecasting method. Note difference in (density) range of the three plots. Values are shown of time
stamps that are both available for the cloud cam and the decorrelation data. The relative high occurrence of zero values in the cloud cam derived CMV origin can be attributed to the
method of using CMV persistence starting at the beginning of the day. Quantitative comparison is presented in Table A.8.

1480
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

Table A.8
Numerical results of the comparison of the different CMV origin methods.

Mean bias error Mean absolute error Median absolute error

De-corr. – (Δ) De-corr. 3.6° 50° 20°


(Δ) De-corr.– Cloud Cam 5.9° 45° 23°
De-corr. – Cloud Cam 0.6° 61° 40°

dij
(Δ) C (d ;a ,b )
ij ij C C = aC + (1−aC )Exp ⎛− ⎞⎜ ⎟

⎝ bC ⎠ (A.7)
The exact value of these parameters is not of interest at this point, as long as the value of aC , that guarantees a fit through = 0) = 1 is (Δ) C (d
ij ij
fixed, such that the de-correlation length bC can be interpreted equally for all of the binned scatter plots at the same moment t and associated TObs .
This offset aC is determined by the average value over all direction bins for distances larger than 20 km and is found to be between 0.3 and 0.6. In
some cases, the scattered values will not decrease quickly enough within the maximal separation of dij ≈ 60 km and the fit model would then balance
optimization of the offset and the de-correlation length bC , which is unwanted.

A.2.1. Cloud motion direction


Instead of using the cloud cam (exogenous input) for determination of the CMV direction for the P2P forecasting method, it can also be inferred
from the cross-correlation between all the PV-systems in the PV-sensor Network at a time t over the observation window range ΔtObs (endogenous
input). The short-term lagged temporal de-correlation fit model (Eq. (A.7)) can be used to assess the direction part of the CMV. The direction bin that
shows the highest value of bC will be interpreted as the direction of origin of the cloud motion as spatially separated locations will be correlated more
strongly along wind, than in the cross-wind direction, see Figs. A.16 for fit examples and A.17 for corresponding processed CMV. The corresponding
instantaneous cloud images with the found motion vectors is shown in Fig. A.14. In situations where the CMV could not be detected, the value of the
previous time step is taken, or otherwise 0° although the concept of a neutral direction is ambiguous. It can occur that there is a false derating of (Δ) Cij
found in an arbitrary direction. A lower threshold of R2 = 0.75 is enforced to exclude such bad fits. Note that the goodness-of-fit R2 is not per se best
for the highest value of bC : at very strong de-correlation, basically all the points are at one line of which the fit is rewarded by a very high R2 as that
value measures the squared error of the fit compared to the squared error of a horizontal line. Fitting the pure scattered data instead of the distance
bin-averaged data would give better differentiation in terms of R2 , but then the maximum value of bC is less clear.
It should be noted that for the mentioned measurement period, the cloud cam images were only available between 5:00 and 14:20 UTC, leading
to uncertainty by assumed persistence of the data beyond that point to cover the studied daytime, see Fig. A.17 (bottom graphs, beyond the red
vertical line). Fig. A.18 shows the cross-correspondence of the three CMV direction origin methods (de-correlation of kt∗, and of Δkt∗ and via the cloud
cam) in the overlapping daily time periods. Generally, there is good agreement between the methods as can be seen from the clustering of values
around the diagonal, although deviations can go in any direction due to occurrences of low amount of available tracking vectors, multiple visible
cloud layers that move in different directions, artifacts (lens-flares/rain droplets) in the case of the cloud cam CMV or in the case of the de-correlation
methods: different correlation maximums between the two methods. From the diagrams it can also be concluded that most of the time during the
measurement period March–August 2015, clouds come in from the −45° to −135° direction (West). The P2P forecasting algorithm uses a selection
window of ± 15°, so minor deviations from the actual CMV, that might be even different from either the cloud cam or the de-correlation methods,
will not result in discarding of the essential highly correlating lagged peripheral time series. The numerical correspondence between the CMV
methods is shown in Table A.8. The Mean Bias Error was calculated as the mean value of all the angle pairs in each method comparison using Eq.
(A.8) for the (vector) difference of two angles (degree).
Δ(θi,θj ) ≡ [θi−θj + 180° Mod 360°]−180° (A.8)

References grid to supporting large-scale renewable energy. J Power Sour 2005;144(1):280–94.


http://dx.doi.org/10.1016/j.jpowsour.2004.12.022.
[8] Bird L, Lew D, Milligan M, Carlini EM, Estanqueiro A, Flynn D, et al. Wind and solar
[1] SPE, 2017. SPE, (Solar Power Europe) global market outlook for solar power energy curtailment: a review of international experience. Renew Sustain Energy Rev
2017–2021; 2017 < http://www.solarpowereurope.org/insights/global-market- 2016;65:577–86. http://dx.doi.org/10.1016/j.rser.2016.06.082.
outllook/ > . [9] Litjens GBMA, Worrell E, Van Sark WGJHM. Influence of demand patterns on the
[2] Kaur A, Nonnenmacher L, Pedro HTC, Coimbra CFM. Benefits of solar forecasting optimal orientation of photovoltaic systems. Solar Energy 2017;155:1002–14.
for energy imbalance markets. Renew Energy 2016;86:819–30. http://dx.doi.org/ http://dx.doi.org/10.1016/j.solener.2017.07.006.
10.1016/j.renene.2015.09.011. [10] 3E. Meta PV - cost-effective integration of Photovoltaics in distribution grids: results
[3] Van der Kam M, Van Sark WGJHM. Smart charging of electric vehicles with pho- and recommendations; 2015 < http://www.metapv.eu/finalreport > .
tovoltaic power and vehicle-to-grid technology in the residential sector: a case [11] Staats M, de Boer-Meulman P, van Sark W. Experimental determination of demand
study. Appl Energy 2015;152:20–30. side management potential of wet appliances in the Netherlands. Sustain Energy,
[4] Verbong GP, Beemsterboer S, Sengers F. Smart grids or smart users? Involving users Grids Netw 2017;9:80–94<http://www.sciencedirect.com/science/article/pii/
in developing a low carbon electricity economy. Energy Policy 2013;52:117–25. S2352467716302144>.
http://dx.doi.org/10.1016/j.enpol.2012.05.003. [12] Fang X, Misra S, Xue G, Yang D. Smart grid - the new and improved power grid: a
[5] Castillo-Cagigal M, Caamano-Martín E, Matallanas E, Masa-Bote D, Gutiérrez A, survey. IEEE Commun Surv Tut 2012;14(4):944–80.
Monasterio-Huelin F, et al. PV self-consumption optimization with storage and ac- [13] la Parra ID, Marcos J, García M, Marroyo L. Storage requirements for PV power
tive DSM for the residential sector. Solar Energy 2011;85(9):2338–48. http://dx. ramp-rate control in a PV fleet. Solar Energy 2015;118:426–40<http://www.
doi.org/10.1016/j.solener.2011.06.028. sciencedirect.com/science/article/pii/S0038092X15003102>.
[6] Matallanas E, Castillo-Cagigal M, Gutiérrez A, Monasterio-Huelin F, Caamaño- [14] DNV-GL, 2015. Power matching city phase II report by DNV-GL http://www.
Martín E, Masa D, et al. Neural network controller for active demand-side man- powermatchingcity.nl/; 2015 < http://www.powermatchingcity.nl/data/docs/
agement with PV energy in the residential sector. Appl Energy 2012;91(1):90–7. PowerMatching%20City_brochure_final_UK_29-04-2015_lowres.pdf > .
http://dx.doi.org/10.1016/j.apenergy.2011.09.004. [15] EBU. Economic board Utrecht - smart solar charging, a unique solution for accel-
[7] Kempton W, Tomić J. Vehicle-to-grid power implementation: from stabilizing the erating the transition to renewable energy - position paper; 2016 < http://www.

1481
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

economicboardutrecht.nl/sites/nl.economicboardutrecht.www/files/Position http://dx.doi.org/10.1016/j.solener.2010.08.011.
%20paper%20smart%20solar%20charging%20-%20EN.pdf > . [44] Voyant C, Muselli M, Paoli C, Nivet M-L. Optimization of an artificial neural net-
[16] Mwasilu F, Justo JJ, Kim E-K, Do TD, Jung J-W. Electric vehicles and smart grid work dedicated to the multivariate forecasting of daily global radiation. Energy
interaction: a review on vehicle to grid and renewable energy sources integration. 2011;36(1):348–59. http://dx.doi.org/10.1016/j.energy.2010.10.032.
Renew Sustain Energy Rev 2014;34:501–16. http://dx.doi.org/10.1016/j.rser. [45] Pedro HTC, Coimbra CFM. Assessment of forecasting techniques for solar power
2014.03.031. production with no exogenous inputs. Solar Energy 2012;86(7):2017–28. http://dx.
[17] Lorenz E, Heinemann D. Earth and planetary sciences, chapter 1.13. Prediction of doi.org/10.1016/j.solener.2012.04.004.
solar irradiance and photovoltaic power. Elsevier Ltd.; 2012. p. 239–92. http://dx. [46] Fonecsa Jr. JG, Oozeki T, Takashima T, Koshimizu G, Uchida Y, Ogimoto K. Use of
doi.org/10.1016/B978-0-08-087872-0.00114-1. ISBN 9780080878737. support vector regression and numerically predicted cloudiness to forecast power
[18] Diagne M, David M, Lauret P, Boland J, Schmutz N. Review of solar irradiance output of a photovoltaic power plant in Kitakyushu, Japan. Prog Photovolt: Res
forecasting methods and a proposition for small-scale insular grids. Renew Sustain Appl 2012;20:874–82. http://dx.doi.org/10.1002/pip.1152.
Energy Rev 2013;27:65–76. http://dx.doi.org/10.1016/j.rser.2013.06.042. [47] Boland J. Spatial-temporal forecasting of solar radiation. Renew Energy
[19] Kleissl J, editor. Solar energy forecasting and resource assessment. Oxford UK: 2015;75:607–16. http://dx.doi.org/10.1016/j.renene.2014.10.035.
Elsevier Academic Press; 2013. ISBN 9780123971777. [48] Rana M, Koprinska I, Agelidis VG. Univariate and multivariate methods for very
[20] Redmund J, Calhau C, Perret L, Marcel D. Characterization of the spatio-temporal short-term solar photovoltaic power forecasting. Energy Convers Manage
variations and ramp rates of solar radiation and PV - Report of IEA Task 14 Subtask 2016;121:380–90. http://dx.doi.org/10.1016/j.enconman.2016.05.025.
1.3. Tech. Rep. IEA PVPS T14-05:2015, International Energy Agency - Photovoltaic [49] Dambreville R, Blanc P, Chanussot J, Boldo D, Dubost S. Very short term forecasting
Power Systems Programme; 2015. of the global horizontal irradiance through Helioclim maps. IREC 2014:291–300.
[21] Antonanzas J, Osorio N, Escobar R, Urraca R, Martinez-de pison FJ, Antonanzas- http://dx.doi.org/10.1016/j.renene.2014.07.012.
torres F. Review of photovoltaic power forecasting. Solar Energy 2016;136:78–111. [50] Zagouras A, Pedro HTC, Coimbra CFM. On the role of lagged exogenous variables
http://dx.doi.org/10.1016/j.solener.2016.06.069. and spatio-temporal correlations in improving the accuracy of solar forecasting
[22] Hammer A, Heinemann D, Lorenz E, Lückehe B. Short-term forecasting of solar methods. Renew Energy 2015;78:203–18. http://dx.doi.org/10.1016/j.renene.
radiation: a statistical approach using satellite data. Solar Energy 2014.12.071.
1999;67(1):139–50. http://dx.doi.org/10.1016/S0038-092X(00)00038-4. [51] Wolff B, Kühnert J, Lorenz E, Kramer O, Heinemann D. Comparing support vector
[23] Pelland S, Galanis G, Kallos G. Solar and photovoltaic forecasting through post- regression for PV power forecasting to a physical modeling approach using mea-
processing of the global environmental multiscale numerical weather prediction surement, numerical weather prediction, and cloud motion data. Solar Energy
model. Prog Photovolt: Res Appl 2013;21(3):284–96. http://dx.doi.org/10.1002/ 2016;135:197–208.
pip.1180. [52] Vaz AGR, Elsinga B, Van Sark WGJHM, Brito MC. An artificial neural network to
[24] Mathiesen P, Collier C, Kleissl J. A high-resolution, cloud-assimilating numerical assess the impact of neighbouring photovoltaic systems in power forecasting in
weather prediction model for solar irradiance forecasting. Solar Energy Utrecht, the Netherlands. Renew Energy 2016;85:631–41<http://linkinghub.
2013;92:47–61. http://dx.doi.org/10.1016/j.solener.2013.02.018. elsevier.com/retrieve/pii/S0960148115300847>.
[25] Mathiesen P, Kleissl J. Evaluation of numerical weather prediction for intra-day [53] Inman RH, Pedro HTC, Coimbra CFM. Solar forecasting methods for renewable
solar forecasting in the continental united states. Solar Energy 2011;85(5):967–77. energy integration. Prog Energy Combust Sci 2013;39(6):535–76. http://dx.doi.
http://dx.doi.org/10.1016/j.solener.2011.02.013. org/10.1016/j.pecs.2013.06.002.
[26] Perez R, Kivalov S, Schlemmer J, Hemker Jr. K, Renné D, Hoff T. Validation of short [54] Widén J, Carpman N, Castellucci V, Lingfors D, Olauson J, Remouit F, et al.
and medium term operational solar radiation forecasts in the {US}. Solar Energy Variability assessment and forecasting of renewables: a review for solar, wind, wave
2010;84(12):2161–72. http://dx.doi.org/10.1016/j.solener.2010.08.014. and tidal resources. Renew Sustain Energy Rev 2015;44:356–75. http://dx.doi.org/
[27] Marquez R, Coimbra CFM. Intra-hour DNI forecasting based on cloud tracking 10.1016/j.rser.2014.12.019.
image analysis. Solar Energy 2013;91:327–36. http://dx.doi.org/10.1016/j.solener. [55] Barbieri F, Rajakaruna S, Ghosh A. Very short-term photovoltaic power forecasting
2012.09.018. with cloud modeling: a review. Renew Sustain Energy Rev 2017;75(August
[28] Chow CW, Urquhart B, Lave M, Dominguez A, Kleissl J, Shields J, et al. Intra-hour 2015):242–63. http://dx.doi.org/10.1016/j.rser.2016.10.068.
forecasting with a total sky imager at the UC San Diego solar energy testbed. Solar [56] Mellit A, Kalogirou Sa. Artificial intelligence techniques for photovoltaic applica-
Energy 2011;85(11):2881–93. http://dx.doi.org/10.1016/j.solener.2011.08.025. tions: a review. Prog Energy Combust Sci 2008;34(5):574–632. http://dx.doi.org/
[29] Chow CW, Belongie S, Kleissl J. Cloud motion and stability estimation for intra-hour 10.1016/j.pecs.2008.01.001.
solar forecasting. Solar Energy 2015;115:645–55<http://www.sciencedirect.com/ [57] Voyant C, Notton G, Kalogirou S, Nivet ML, Paoli C, Motte F, et al. Machine learning
science/article/pii/S0038092X15001565>. methods for solar radiation forecasting: a review. Renew Energy 2017;105:569–82.
[30] Chu Y, Pedro HTC, Coimbra CFM. Hybrid intra-hour DNI forecasts with sky image http://dx.doi.org/10.1016/j.renene.2016.12.095.
processing enhanced by stochastic learning. Solar Energy [58] Murata A, Otani K. An analysis of time-dependent spatial distribution of output
2013;98:592–603<http://linkinghub.elsevier.com/retrieve/pii/ power from very many PV power systems installed on a nation-wide scale in Japan.
S0038092X13004325>. Solar Energy Materials Solar Cells 1997;47(1–4):197–202. http://dx.doi.org/10.
[31] Chu Y, Pedro HTC, Li M, Coimbra CFM. Real-time forecasting of solar irradiance 1016/S0140- 6701(97)82039-5.
ramps with smart image processing. Solar Energy 2015;114:91–104<http:// [59] Wiemken E, Beyer H, Heydenreich W, Kiefer K. Power characteristics of PV en-
linkinghub.elsevier.com/retrieve/pii/S0038092X15000389>. sembles: experiences from the combined power production of 100 grid connected
[32] Peng Z, Yu D, Huang D, Heiser J, Yoo S, Kalb P. 3D cloud detection and tracking PV systems distributed over the area of Germany. Solar Energy 2001;70(6):513–8.
system for solar forecast using multiple sky imagers. Solar Energy http://dx.doi.org/10.1016/S0038-092X(00)00146-8.
2015;118(August):496–519. http://dx.doi.org/10.1016/j.solener.2015.05.037. [60] Perez R, Kivalov S, Schlemmer J, Hemker K, Hoff TE. Short-term irradiance
[33] Muselli M, Poggi P, Notton G, Louche A. First order markov chain model for gen- variability: preliminary estimation of station pair correlation as a function of dis-
erating synthetic “typical days” series of global irradiation in order to design tance. Solar Energy 2012;86(8):2170–6. http://dx.doi.org/10.1016/j.solener.2012.
photovoltaic stand alone systems. Energy Convers Manage 2001;42(6):675–87. 02.027.
http://dx.doi.org/10.1016/S0196-8904(00)00090-X. [61] Hoff TE, Perez R. Modeling PV fleet output variability. Solar Energy
[34] Li Y-Z, Luan R, Niu J-c. Forecast of power generation for grid-connected photo- 2012;86(8):2177–89. http://dx.doi.org/10.1016/j.solener.2011.11.005.
voltaic system based on grey model and markov chain. 3rd IEEE conference on [62] Perez MJR, Fthenakis VM. On the spatial decorrelation of stochastic solar resource
industrial electronics and applications, 2008, ICIEA 2008 IEEE; 2008. p. 1729–33. variability at long timescales. Solar Energy 2015;117:46–58. http://dx.doi.org/10.
http://dx.doi.org/10.1109/ICIEA.2008.4582816. 1016/j.solener.2015.04.020.
[35] Duda, Hart. Pattern classification. 2nd ed. New York: John Wiley & Sons; 2000. [63] Lave M, Kleissl J, Arias-Castro E. High-frequency irradiance fluctuations and geo-
[36] Yakowitz S. Nearest-neighbour methods for time series analysis. J Time Ser Anal graphic smoothing. Solar Energy 2012;86(8):2190–9<http://linkinghub.elsevier.
1987;8(2):235–47. http://dx.doi.org/10.1111/j.1467-9892.1987.tb00435.x. com/retrieve/pii/S0038092X11002611>.
[37] Reikard G. Predicting solar radiation at high resolutions: a comparison of time [64] Hinkelman LM. Differences between along-wind and cross-wind solar irradiance
series forecasts. Solar Energy 2009;83(3):342–9<http://linkinghub.elsevier.com/ variability on small spatial scales. Solar Energy 2013;88:192–203. http://dx.doi.
retrieve/pii/S0038092X08002107>. org/10.1016/j.solener.2012.11.011.
[38] Bacher P, Madsen H, Nielsen HA. Online short-term solar power forecasting. Solar [65] Perpiñán O, Marcos J, Lorenzo E. Electrical power fluctuations in a network of DC/
Energy 2009;83(10):1772–83. http://dx.doi.org/10.1016/j.solener.2009.05.016. AC inverters in a large PV plant: relationship between correlation, distance and time
[39] Yang D, Sharma V, Ye Z, Lim LI, Zhao L, Aryaputera AW. Forecasting of global scale. Solar Energy 2013;88:227–41.
horizontal irradiance by exponential smoothing, using decompositions. Energy [66] Elsinga B, Van Sark WGJHM. Spatial power fluctuation correlations in urban
2015;81:111–9<http://linkinghub.elsevier.com/retrieve/pii/ rooftop photovoltaic systems. Prog Photovolt: Res Appl 2015;23(10):1390–7.
S0360544214013528>. http://dx.doi.org/10.1002/pip.2539.
[40] Hyndman R, Koehler AB, Ord JK, Snyder RD. Forecasting with exponential [67] Lohmann GM, Monahan AH, Heinemann D. Local short-term variability in solar
smoothing: the state space approach. Springer Science & Business Media; 2008. irradiance. Atmos Chem Phys 2016;16(10):6365–79. http://dx.doi.org/10.5194/
[41] Urraca R, Antonanzas J. Smart baseline models for solar irradiation forecasting acp-16-6365-2016.
Spanish agency for irrigation in agriculture. Energy Convers Manage [68] Elsinga B, Van Sark WGJHM. Inter-system time lag due to clouds in an urban PV
2016;108:539–48. http://dx.doi.org/10.1016/j.enconman.2015.11.033. ensemble. In: Conference record of the IEEE photovoltaic specialists conference;
[42] Sfetsos A, Coonick A. Univariate and multivariate forecasting of hourly solar ra- 2014. p. 754–8.
diation with artificial intelligence techniques. Solar Energy 2000;68(2):169–78. [69] Lonij VP, Brooks AE, Cronin AD, Leuthold M, Koch K. Intra-hour forecasts of solar
http://dx.doi.org/10.1016/S0038-092X(99)00064-X. power production using measurements from a network of irradiance sensors. Solar
[43] Paoli C, Voyant C, Muselli M, Nivet M-L. Forecasting of preprocessed daily solar Energy 2013;97:58–66. http://dx.doi.org/10.1016/j.solener.2013.08.002.
radiation time series using neural networks. Solar Energy 2010;84(12):2146–60. [70] Lorenzo AT, Holmgren WF, Cronin AD. Irradiance forecasts based on an irradiance

1482
B. Elsinga, W.G.J.H.M. van Sark Applied Energy 206 (2017) 1464–1483

monitoring network, cloud motion, and spatial averaging. Solar Energy Dec 2016; 2016 < http://uppenergy.com/product/details/1f > .
2015;122:1158–69. http://dx.doi.org/10.1016/j.solener.2015.10.038. [82] Orgill J, Hollands K. Correlation equation for hourly diffuse radiation on a hor-
[71] Yang D, Ye Z, Lim LHI, Dong Z. Very short term irradiance forecasting using the izontal surface. Solar Energy 1977;19:357–9. http://dx.doi.org/10.1016/0038-
lasso. Solar Energy 2015;114:314–26<http://linkinghub.elsevier.com/retrieve/ 092X(77)90006-8.
pii/S0038092X15000304>. [83] Perez RR, Ineichen P, Steward RDM. A new simplified version of the Perez diffuse
[72] Yang D, Gu C, Dong Z, Jirutitijaroen P, Chen N, Walsh WM, et al. Solar irradiance irradiance model for tilted surfaces. Solar Energy 1987;39:221–31.
forecasting using spatial-temporal covariance structures and time-forward kriging. [84] Wolfram Research Inc.. Mathematica Version 11. Champaign (IL); 2016.
Renew Energy 2013;60:235–45<http://linkinghub.elsevier.com/retrieve/pii/ [85] Ineichen P, Perez R. A new airmass independent formulation for the linke turbidity
S0960148113002759>. coefficient. Solar Energy 2002;73(3):151–7. http://dx.doi.org/10.1016/S0038-
[73] Yang D, Dong Z, Reindl T, Jirutitijaroen P, Walsh WM. Solar irradiance forecasting 092X(02)00045-2.
using spatio-temporal empirical kriging and vector autoregressive models with [86] Van Sark WGJHM, Louwen A, de Waal AC, Elsinga B, Schropp REI. UPOT: the
parameter shrinkage. Solar Energy 2014;103:550–62. http://dx.doi.org/10.1016/j. Utrecht photovoltaic outdoor test facility. In: 27th EU PVSEC; 2012. p. 3247–9.
solener.2014.01.024. [87] Louwen A, de Waal AC, Schropp REI, Faaij APC, van Sark WGJHM. Comprehensive
[74] Aryaputera AW, Yang D, Zhao L, Walsh WM. Very short-term irradiance forecasting characterisation and analysis of PV module performance under real operating
at unobserved locations using spatio-temporal kriging. Solar Energy conditions. Prog Photovolt: Res Appl 2016;25:218–32. http://dx.doi.org/10.1002/
2015;122:1266–78<http://linkinghub.elsevier.com/retrieve/pii/ pip.2848. pp. 2848.
S0038092X15005745>. [88] Marquez R, Coimbra CFM. Proposed metric for evaluation of solar forecasting
[75] Cressie N, Wikle CK. Statistics for spatio-temporal data. New York: John models. J Solar Energy Eng 2013;135(1):011016.
Wiley & Sons; 2011. ISBN 9780471692744. [89] Stein JS, Hansen CW, Reno MJ. The variability index: a new and novel metric for
[76] Achleitner S, Kamthe A, Liu T, Cerpa AE. Sips: solar irradiance prediction system. quantifying irradiance and PV output variability In: Sandia National Laboratories,
In: Proceedings of the 13th international symposium on information processing in Albuquerque (NM); 2012 < http://energy.sandia.gov/wp/wp-content/gallery/
sensor networks, IPSN-14. IEEE; 2014. p. 225–36. uploads/Stein_ASES_2012_VI_paper_SAND2012-2088C1.pdf > .
[77] Lipperheide M, Bosch J, Kleissl J. Embedded nowcasting method using cloud speed [90] Mazorra Aguiar L, Pereira B, David M, Díaz F, Lauret P. Use of satellite data to
persistence for a photovoltaic power plant. Solar Energy 2015;112:232–8. http:// improve solar radiation forecasting with Bayesian Artificial Neural Networks. Solar
dx.doi.org/10.1016/j.solener.2014.11.013. Energy 2015;122:1309–24<http://linkinghub.elsevier.com/retrieve/pii/
[78] Bosch JL, Kleissl J. Cloud motion vectors from a network of ground sensors in a S0038092X15005927>.
solar power plant. Solar Energy 2013;95:13–20. http://dx.doi.org/10.1016/j. [91] Hammer A, S K. Analyse kurzfristiger fluktuationen der solarstrahlung unter
solener.2013.05.027. berücksichtigung von wolkenfeldstrukturen, diploma thesis; 1993.
[79] Fung V, Bosch JL, Roberts SW, Kleissl J. Cloud shadow speed sensor. Atmos Measur [92] Kannala J, Brandt S. A generic camera model and calibration method for conven-
Techn 2014;7(6):1693–700<http://www.atmos-meas-tech.net/7/1693/2014/>. tional, wide-angle, and fish-eye lenses. IEEE Trans Pattern Anal Mach Intell
[80] Elsinga B, Van Sark WGJHM, Ramaekers LA. Inverse photovoltaic yield model for 2006:28. http://dx.doi.org/10.1109/TPAMI.2006.153.
global horizontal irradiance reconstruction. Energy Sci Eng 2017;5(4):226–39. [93] Urquhart B, Kurtz B, Kleissl J. Sky camera geometric calibration using solar ob-
http://dx.doi.org/10.1002/ese3.162. servations. Atmos Measur Techn 2016;9:4279–94. http://dx.doi.org/10.5194/amt-
[81] Upp Energy. Upp energy (part of Agile XS), sensor product 1f. Last referenced on 20 9-4279-2016.

1483

You might also like