Академический Документы
Профессиональный Документы
Культура Документы
of Intermittent Power Sources
Zongjie Wang 1,* and Zhizhong Guo 1, 2
1 School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin 150001, China
fylx‐001@163.com (W.Z.); zhizhonggzz@sina.com (G.Z.)
2 Harbin Institute of Technology at Zhangjiakou ITRIZ, Zhangjiakou 075400, China; zhizhonggzz@sina.com
(G.Z.)
* Correspondence: fylx‐001@163.com; Tel.: +86‐189‐1029‐4513
Abstract: This paper establishes a statistical quantification of the uncertainty levels of intermittent
power sources. We first construct a negative exponential function, referred to as a statistical
function, to represent the relationship between the statistical regularity of the forecast error of a
single intermittent power source and the time ahead of the forecast. Subsequently, we generalize
this negative exponential function to a family of statistical functions, namely the sum statistical
functions, the equivalent statistical functions, and the contour statistical functions, which are
proposed to characterize the overall statistical forecast uncertainty levels of multiple intermittent
power sources and all power sources. Based on historical observations, parameters of these
functions are estimated to represent the statistical regularity of the forecast uncertainty levels of all
the power sources of interest. Historical data sampled from real wind farms and solar power sites
demonstrate the effectiveness of the proposed method.
Keywords: intermittent power sources; statistical regularity; power generation uncertainty level;
forecast error
1. Introduction
Intermittent Power Sources (IPSs), such as wind and solar power generators, have been
extensively integrated into power systems1‐6. However, power generation from these sources is highly
unpredictable due to their intermittency, randomness, and volatility of the energy supply7‐11. The
overall uncertainty of a power system depends on the uncertainty of each IPS and the proportion of
IPS generation within the total energy generation12‐16. In 2014, a report from the National Renewable
Energy Laboratory predicted that renewable electricity generation would be more than adequate to
supply 80% of the total electricity generation in the USA in 205017. However, such a large amount of
unpredictable power generation may jeopardize the security and stability of power systems. Indeed,
the uncertainty of IPSs has been widely acknowledged as one of the most significant challenges to
real‐time power balancing of load and supply18‐23.
The main source of the uncertainty of IPSs is the unpredictability due to forecasting the power
available for the “time ahead”, such as 1 hour ahead or 24 hours ahead. This can be represented by a
forecast error. Several algorithms for day‐ahead forecasts of IPS generation have been developed24‐29.
However, to the best of our knowledge, obtaining satisfactory forecast accuracy for IPS generation is
still an unsolved problem. The intrinsic reason is that the 24‐hour ahead forecast that is traditionally
utilized for other non‐IPSs is not useful for IPSs, because it is too long into the future to produce an
accurate forecast. In an attempt to address this, the most direct method is to shorten the time ahead
of forecast for IPSs. However, this imposes higher requirements on the efficiencies of dispatch and
market operations. To determine the best time ahead of forecast for IPSs, a quantitative and statistical
evaluation of the overall uncertainty of all power sources is highly desirable. Specifically, we are
interested in the temporal characteristics of the statistical forecast uncertainty level.
There have been many publications investigating the statistical regularity of power generation
uncertainty30‐37. In reference [30], Xu et al., employed a maximum entropy spectrum estimation
2 of 19
method to study the temporal and spatial distribution characteristics of wind power periodicity. In
reference [31], Sideratos et al., presented an advanced statistical method for wind power forecast
based on artificial intelligence technologies. In reference [32], Li et al., studied the statistical regularity
of wind power fluctuation in different spatial‐temporal scales from the aspects of smoothness and
correlation. Even more closely related to the subject of this paper, references [33–37] presented
statistical studies of the relationship between the forecast error of IPS generation and the time ahead
of forecast. For example, references [35] and [36] show that as the time ahead increases, the forecast
error increases quickly initially, but then grows much more slowly after a certain point (called point
b). Based on the statistical regularity, the randomness of individual examples was eliminated, and a
relationship between the uncertainty and time ahead of forecast was established.
Statistical quantification of the uncertainty level of multiple IPSs and all power sources is still an
unsolved problem. Unlike the existing literature reviewed above, this paper focuses on analytical
formulas for the statistical regularity of forecast uncertainty levels of IPSs with respect to time ahead.
These formulas quantitatively characterize the uncertainty levels of power generation from wind or
solar power sources. With regards to the relationship between the forecast error of IPS generation
and the time ahead of forecast, our objective is to extract the statistical regularity of the uncertainty
level of IPS generation by quantitatively characterizing a large number of historical observations by
using the statistical functions derived in this paper.
We first define the relative Root Mean Square Error (RMSE) of IPSs as the statistical forecast
uncertainty level of IPSs, and then employ a family of negative exponential functions to represent the
temporal characteristics of the statistical forecast uncertainty level. Specifically, we define: statistical
functions of the forecast error of a single IPS, the sum statistical function of multiple IPSs as well as
all power sources that can be represented by the equivalent statistical function with a variant scaling
factor, and the contour statistical function, to portray a closure of the sum statistical function. We
compare these functions with a large amount of real data from references [36] and [37], and present
a method for estimating the parameters of these statistical functions. The functions can succinctly
represent the temporal trend of the statistical forecast uncertainty level of all power sources.
Comprehensive assessment of the IPS quality can be achieved by considering statistical functions
with other factors such as wind speed, wind direction, irradiation time, and irradiation intensity.
Moreover, the proposed statistical functions can provide quantitative characterization of the
intermittent power uncertainty level corresponding to different methodologies of power network
planning and operation with a high proportion of IPS generation, such as power system security
analysis, reliability assessment, and economical operation.
The remainder of this paper is organized as follows. Based on statistical regularity, we first
propose a statistical function of the forecast error in Section 2. The uncertainty of IPSs with statistical
functions is quantified in Section 3. The uncertainty of all power sources with statistical functions is
further quantified in Section 4. The simulation results are demonstrated in Section 5. Section 6
discusses some properties and applications, and conclusions are finally drawn in Section 7.
2. Quantifying the statistical forecast uncertainty level of wind or solar power sources
2.1 Statistical regularity
Based on a large amount of real on‐site data, previous studies have established the statistical
relationship between the RMSE of wind and solar power source forecasts and the time ahead of day‐
ahead scheduling and look‐ahead dispatch34‐37. In this paper, such a relationship is defined as the
uncertainty level of IPS generation, and is referred to as the rmse curve.
3 of 19
Figure 1. rmse
w curve of a wind farm in northwestern USA [36].
Figure 2. rmse
s curve of a photovoltaic power plant in northern France [37].
negative exponential function, , i.e., a statistical function that is consistent with the general rule of
the prediction of future events, is proposed to describe the statistical regularity of the forecast
uncertainty level of IPS generation.
2.2 Statistical functions of forecast error
In this subsection, we first define the Statistical Function of Forecast Error (SFFE) of IPSs to
characterize the statistical forecast uncertainty level of a single IPS with respect to the time ahead t,
represented by t , which is shown as follows
t
t (1 e ), (t [0, )) . (1)
Herein, the amplitude (0, ) and the time coefficient (0, ) satisfy the following
boundary conditions
, (2)
(1) (0) (1) (0)
where (1) (0) is the first derivative of t with respect to t when t is 0.
To better demonstrate the influence of the statistical forecast uncertainty level with respect to
the time ahead t, we normalize t in equation (1) and define the following Normalized Statistical
Function of Forecast Error (NSFFE), which is represented by t as follows
t
t
t 1 e . (3)
Next, we establish some important properties for the SFFE and the NSFFE, i.e., the function
and the function.
Proposition 1. The NSFFE ( function) yields the following conservation principle
1, (i 1)
(i ) (i 1) = , (4)
0, (i 2)
where (i ) is the ith‐order derivative of the function.
Proof. Please see the detailed proof in the Appendix.
Corollary 1. The SFFE ( function) yields the following conservation principle
( i ) (t ) ( i 1) (t ) 1, (i 1)
. (5)
(1) (0) 0, (i 2)
Proof. Please see the detailed proof in the Appendix.
In real‐world terms, parameters and in equation (1) can be interpreted as follows.
The amplitude represents the statistical forecast error of IPS generation when the time ahead
tends to infinity. Assuming that different forecast techniques converge to similar performances when
the time ahead tends to infinity, the parameter actually describes the objective uncertainty level
of IPS generation. In practice, the statistical forecast error of IPSs usually approaches its amplitude
with a time ahead of around 24 hours. Thus, parameter can be approximated by the day‐ahead
statistical forecast error of IPSs.
The time coefficient defines the growth rate of t when t = 0. It represents how advanced
the forecast technique is in leveraging temporal correlations to better infer future IPS generation, i.e.,
to reduce the growth rate of the uncertainty level with respect to the time ahead. A greater
indicates a slower growth rate of the uncertainty level of IPSs, i.e., a better credibility of the forecast.
Consequently, t represents a quantitative characterization of the statistical forecast uncertainty
level of IPSs as a function of the time ahead t.
2.3 Statistical functions of the forecast errors of wind or solar power sources
5 of 19
Figure 3. Time‐ahead curves and actual curve for daily samples.
6 of 19
3. Statistical forecast uncertainty levels
3.1 SSFs
In order to characterize the sum statistical forecast uncertainty level of all IPSs, we define the
SSF with respect to the time ahead t, represented by t and shown as follows
n
(t ) (t ) , (20)
i 1
i
where n represents the quantity of functions.
Let
7 of 19
n
(21)
i 1
i
and
i
i , (i 1, 2,..., n ) , (22)
where i represents the distribution coefficient.
then
t
n
(t ) i (1 e ) . (23)
i
i 1
Similarly, we normalize the t in equation (20) and define the following Normalized Sum
Statistical Function (NSSF), which is represented by t as follows
(t ) n
t
(t ) = i (1 e ) . (24)
i
i 1
Consequently, equation (23) can be transformed into
(t ) (t ) . (25)
3.2 Equivalent statistical functions
The (t ) in equation (20) is the sum of multiple negative exponential SFFEs. For simplicity, in
this subsection, we denote the SSF by a single negative exponential function with a variant time
coefficient (t ) . Such a function is referred to as the Equivalent Statistical Function (ESF), represented
by (t , (t )) and shown as follows
t
(t , (t )) (1 e (t )
) , (26)
where the time coefficient (t ) is a function of the time ahead t.
For t 0, and n given SFFE ( functions), there exists a (t ) such that the following
equation holds
n
(t , (t )) (t ) (t ) , (27)
i 1
i
where
0 (0) . (31)
Proof. Please see the detailed proof in the Appendix.
To better demonstrate the properties established above, we depict a possible group of curves in
Figure 4 as an example. In this example, we assume there are two NSFFEs ( (t ) functions), say 1 (t )
8 of 19
i 1
According to the first term in equation (30), the initial value of (t ) is
0.8 0.2 1
(0) 0 ( ) 3.33 . (33)
4 2
(a) (t ) (b)
Figure 4. Examples of the (t ) and functions.
The curves in Figure 4(a) show that, in this example, the time coefficient function (t ) is indeed
monotonically increasing. The initial value (0) is equal to 0 3.33 and () converges to
1 4 . The curves in Figure 4(b) indicate that the curve of the NESF ( (t , (t )) function) matches the
curve of the NSSF ( (t ) function). Hence, these observations are consistent with the propositions
established in this subsection.
3.3 Contour statistical functions
Given a group of SFFEs ( (t ) ) with 1 , 2 ,..., n , we define the Contour Statistical Function
(CSF) as follows
t
0
(t , 0 ) (1 e ) . (34)
The corresponding function is referred to as the Normalized Contour Statistical Function
(NCSF), i.e.,
t
(t , 0 ) 1 e . (35)
0
i 1 i
Proof. Please see the detailed proof in the Appendix.
9 of 19
where
1 1 1
( )i . (39)
i 0
This indicates that the difference between the NESF and the NCSF, i.e., (t ) , is related to the
difference between the reciprocal of the time coefficients. A smaller differential value can be helpful
in reducing (t ) .
The CSF defined in this subsection provides an upper bound for the ESF. Specifically, the CSF is
equal to the ESF at points t 0 and t , respectively. At other points, the CSF lies strictly above
the ESF.
Since the statistical functions are proposed to quantitatively characterize the statistical forecast
uncertainty level, the statistical functions are next employed in real power systems to quantify the
statistical uncertainty level of multiple IPSs (combinations of wind and solar power sources) and all
power sources (combination of intermittent and controllable power sources).
4. Quantifying statistical forecast uncertainty levels of power system sources
4.1 Statistical forecast uncertainty level of IPSs
4.1.1 SSFs of IPSs
In this paper, it is assumed that IPSs consist of wind and solar power sources. Let Pw and Ps
be, respectively, the power generation of the wind and solar power sources, and then we merge them
into the overall power generations from IPSs, say Pips , i.e.,
Pips (t ) Pw (t ) Ps (t ) . (40)
The proportion of the power generation Pw of wind farms in the overall IPS generation Pips is
represented by w , which is defined as
Pw (t )
w (t ) , ( w (t ) [0,1]) . (41)
Pips (t )
The proportion of the solar power is
Ps (t )
1 w (t ) . (42)
Pips (t )
As described in Section 3, here we define ips (t ) as the SSF of the IPS (which is (t ) ) to
represent the overall statistical uncertainty level of IPSs. Consequently, equation (20) becomes
ips (t ) w (t ) w (t ) (1 w (t )) s (t ) . (43)
In this paper, w (t ) represents the proportion of the wind power generation in IPS generation
when the time ahead is zero. Hence, we assume that w (t ) w (0) w , and then the SSF of the IPS (
ips (t ) function) can be simplified as
ips (t ) w w (t ) (1 w ) s (t ) . (44)
The equation above is equivalent to
10 of 19
4.1.2 ESFs of IPSs
The ESF that is equivalent to the SSF ( ips (t ) function) is as follows
t
( t , ( t )) ips ( t ) ips (1 e (t )
) , (49)
where ips denotes the amplitude of the SFFE of the IPS.
The NESF that corresponds to the ESF ( (t , (t )) function) is as follows
t
(t , (t )) ips (t ) 1 e (t )
. (50)
4.1.3 CSFs of IPSs
Furthermore, we employ the CSF of the IPS to establish a closure of the ESF. The CSF ( (t , 0 )
function) is
t
0
(t , 0 ) ips (1 e ) , (51)
where
1 1
0 (0) ( ) . (52)
w s
The NCSF ( (t , 0 ) function) is expressed as
1
ips (t, 0 ) ips (t, 0 ) . (53)
ips
4.2 Statistical forecast uncertainty level of all power sources
Besides IPSs, the power system also comprises controllable generation assets such as thermal
power, hydropower, nuclear power, and energy storage facilities. If the power generation of the
controllable generation assets is Pc , then the overall power generation from the power system, Pg ,
can be expressed as follows
Pg (t ) Pc (t ) Pips (t ) . (54)
Herein, the IPS power proportion Pips (t ) relative to Pg (t ) is referred to as ips ,
Pips (t )
ips (t ) , ( ips (t ) [0,1]) . (55)
Pg (t )
In this paper, ips (t ) represents the proportion of the IPS generation in the overall power
generation when the time ahead is zero. Here, we assume that ips (t ) is a constant ips . We define
g (t ) as the SSF of all power sources ( ) to represent the overall statistical uncertainty level of all
power sources. Hence
g (t ) ips ips (t ) . (56)
The ESF of all power sources ( g (t , (t )) function) is
11 of 19
t
g ( t , ( t )) g (1 e (t )
) g g ( t , ( t )) , (57)
where g is the amplitude of the SFFE of all power sources, and satisfies
g ips ips ips w w ( ips ips w ) s . (58)
Note that in equation (57), the NESF ( g (t , (t )) function) that corresponds to the SSFs of all
power sources g (t , (t )) is the same as that of the IPS. This indicates that ips has nothing to do
with the time coefficient function, i.e., the time coefficient function exhibits “inheritance”. In equation
(58), the amplitude g of the SSF of all power sources is ips times less than the amplitude ips of
the IPS, which exhibits “dilution”.
The CSF of all power sources is
g ( t , 0 ) g ( t , 0 ) . (59)
From the perspective of power systems, the SFFE of the wind or solar power source ( w
function or s function) and the SFFE of the IPS ( ips function) will ultimately be merged as the
SFFE of all power sources ( g function), which quantifies the statistical uncertainty level in the
overall power generation. This function provides an important reference for many power system
applications such as real‐time power balancing, security analysis, and electric power quality control.
5. Examples of statistical functions of power systems with IPSs
5.1 Uncertainty levels of wind or solar power sources
Taking a real wind power farm in northwestern USA[36] and a real solar power park[37] as
examples, the SFFEs of the wind or solar power sources ( function) can be determined based on
the curves of wrmse and srmse shown in Figures 1 and 2, respectively. According to equations (12)
and (13), the resulting SFFEs can be obtained as follows
t
Figure 5. curves and functions of wind and solar power sources.
rmse
12 of 19
5.2 Uncertainty levels of IPSs
For a power system with an IPS, the SFFE of the wind and solar power sources ( w function
and s function), are shown in equation (60). Set w to 80%. From equation (44), the SSF of the IPS (
ips (t ) function) is expressed as
t t
ips (t ) 25.49 (1 e 2.67
) 8.38 (1 e 0.89
) (61)
and
ips 0.8w 0.2 s 33.87 % . (62)
The coefficient is
w
w 0.75 . (63)
ips
The time coefficient 0 of the CSF of the IPS can be obtained as
1 1
0 ( ) 1.79 h . (64)
w s
Therefore, the CSF of the IPS is
t
ips (t , 0 ) 33.87 (1 e 1.79
) % . (65)
Figure 6 shows that the CSF of the IPS “adheres” to the SSF of the IPS “from the outside towards
the inside”, and the maximum deviation between the two is 2.24% (as denoted by “*” in the figure).
This indicates that the CSF of the IPS can well depict the statistical uncertainty level of the IPS.
5.3 Uncertainty levels of all power sources
For a power system consisting of an IPS and controllable generators, the corresponding SSF of
the IPS is shown in equation (61). Set ips to 60%. Then, based on equation (58), the CSF of all power
sources is
t
g (t , 0 ) 20.32 (1 e 1.79
) % . (66)
Figure 6. SSFs and CSFs of IPSs.
6. Discussion
6.1 Statistical regularity uncertainty levels
13 of 19
To quantitatively characterize the statistical forecast uncertainty of a single IPS or multiple IPSs
or all power sources, we consider the long‐term uncertainty level of IPS generation. By long‐term, we
mean the statistical regularity of forecast errors based on a large number of historical observations.
Note that the statistical forecast uncertainty represents the average forecast error in a wide time
window. Our work is relevant to, but different from, the assessment of short‐term uncertainty, in
which as much information as possible is utilized to infer the output of IPSs in 15 minutes or the
following day. Short‐term uncertainty analysis is a crucial function in real‐time dispatch, look‐ahead
dispatch, and day‐ahead scheduling. For comparison, the long‐term uncertainty studied in this paper
lends itself to mid‐ or long‐term operations and planning of power systems to be able to deal with
every possible individual example that might occur; hence, it should be incorporated into the
statistical regularity of long‐term uncertainty. In particular, the SSF represents the overall statistical
uncertainty of a power system with an IPS, which is an important indicator of predictability.
6.2 ‘Proximity’ property of statistical functions
Statistical functions ( functions) have the ’proximity’ property. First, the amplitude condition
and the time coefficient condition guarantee that functions can include all points underneath the
rmse curve. Second, Proposition 4 guarantees that the CSF includes all points underneath the SSF.
The ’proximity’ property represents conservativeness, namely, all points on the rmse curve are
guaranteed to be no larger than those in the functions. Similarly, all points in the SSF are
guaranteed to be no larger than those in the CSF. In other words, the function characterizes an
upper bound of the uncertainty level estimated from a large number of historical observations. Such
conservativeness is necessary for long‐term planning of power systems.
Conservativeness has somehow exaggerated the statistical forecast uncertainty level. For power
systems with IPSs, the statistical uncertainty level would rather be appropriately exaggerated to a
certain degree to deal with power systems with larger uncertainties, so that power systems with
smaller uncertainties can also be dealt with.
6.3 Applications
As what we discussed in the Introduction section, the proposed statistical functions can be
utilized not only in improving IPS quality and providing a theoretical characterization of IPS
uncertainty level to different power network methods, but also in important practical applications
such as on adjusting the dispatch control system and designing the real‐time dispatch, which are
listed as detailed below.
6.3.1 Real‐time dispatch timescale
The hierarchy of power system operation typically comprises short‐term scheduling, real‐time
dispatch, and automatic generation control (AGC). Due to the integration of intermittent power
sources, the increasing penetration of IPSs may result in larger discrepancies between the AGC and
the actual power flow. The AGC is the last‐level defense of the systemwide power balance and has
limited power adjustment capability, which makes it difficult for the AGC to address the rest of
power imbalance. Therefore, from the perspective of power balance, real‐time dispatch in power
systems with a high penetration of IPSs should be considered and employed in the process of power
balance, to eliminate as much power imbalance as possible from short‐term scheduling. The
capability of doing so depends on the timescale.
Many publications have acknowledged the importance of timescale design in the power
balancing architecture, offering suggestions on possible timescales of real‐time dispatch such as 4
hours, 1 hour, and 15 minutes38‐42. Nevertheless, neither analytical formulas nor theoretical or
systematic studies have yet been proposed on the timescale of real‐time dispatch. This could lead to
two consequences. First, the inappropriate timescale of real‐time dispatch might make AGC difficult
to balance the rest of power imbalance. Second, with the decrease of the real‐time dispatch timescale,
more accurate real‐time forecast of IPS power can be obtained, whereas the computation burden
would be exacerbated as well.
14 of 19
Adequate power sources to accommodate the uncertainty level of IPS and proper design of
timescales are the fundamental insurances for the real‐time dispatch to achieve the best possible
function of power balancing. Therefore, it is highly desirable to design a proper timescale, which
essentially represents the tradeoff between the forecast accuracy and the computation efficiency.
Based on the statistical functions proposed in this paper, the parameters of the timescale of real‐
time dispatch with a high proportion of IPS generation can be determined. Combined with the power
balance capability of AGC, a quantified relationship between the uncertainty level of the power
system and the critical timescale for real‐time dispatch has been further established and applied to
dispatching systems. This application has been particularly extended and accomplished in paper [43].
We briefly introduce how to design the real‐time dispatch timescale as follows.
To satisfy the AGC constraint, that is, to balance the actual power imbalance from the real‐time
dispatch level, by utilizing the CSF of all power sources g (t ) , we obtain the following inequality
which explicitly characterizes the restriction on the timescale of real‐time dispatch imposed by AGC
limits
t
g
g 24 (1 e ) g (t ) A . (67)
where A represents the AGC power adjustment capability range of a power system. Generally, it
is not greater than 10% of the maximum system power generation, here we assume it as 8%.
When the equality holds in (67), the corresponding timescale is referred to as the “critical” real‐
time dispatch timescale, denoted by Tcrr
1
Tcrr g ln , (68)
1
where Tcrr represents the critical timescale of real‐time dispatch, the ratio is defined as
A A
. (69)
g 24 ips ips 24
The selection of the real‐time dispatch timescale TR needs to satisfy two conditions
TR Tcrr
. (70)
kTR 24( h), ( k is a positive integer )
The first formula is easy to understand; the second formula is to connect with dispatching period
of short‐term (day‐ahead) scheduling.
6.3.2 Comparative Case Study
A comparative case study is presented to show the effectiveness of this application on designing
the real‐time dispatch timescale.
The real power system in Zhangjiakou, which is a zonal network with the highest penetration
level of IPS ( ips =60.82%) in China, is tested. Currently, its real‐time dispatch timescale is 1 h and the
AGC power adjustment capability range is 8%.
From a large amount of observations over 1 year, the CSF of IPS ips , and the CSF of all power
sources g , can be obtained as follows
t
ips (t ) 40.87 (1 e 2.61
) (%) , (71)
t
g (t ) 24.86 (1 e 2.61
) (%) . (72)
The critical real‐time dispatch timescale can be calculated according to (68) and (69) as
1 1
Tcrr g ln =2.61 ln 1.01( h ) . (73)
1 1
8
24.86
According to (70), the real‐time dispatch timescale TR can be designed as 1h. This result is
consistent with the current real‐time dispatch timescale. Our simulation indicates that the current
real‐time dispatch timescale is suitable from the perspective of power balancing.
15 of 19
However, even more IPS will be integrated into the power system in the future. According to
[17], a report from the National Renewable Energy Laboratory predicted that renewable electricity
generation would be more than adequate to supply 87% of the total electricity generation in the global
in 2050. Here we assume the ips of Zhangjiakou power network will be increased to 87%, then the
CSF of all power sources will be
t
g (t ) 35.56 (1 e 2.61
) (%) . (74)
The critical real‐time dispatch timescale will be:
1 1
Tcrr g ln =2.61 ln 0.66( h ) . (75)
1 1
8
35.56
The result is smaller than the current real‐time dispatch timescale. This indicates that the current
real‐time dispatch timescale needs to be shortened in the future to accommodate more power
generations from IPS.
According to equation (70), the new real‐time dispatch timescale TR can be properly designed as
0.6 h. Therefore, adequate power sources can be utilized to accommodate the IPS uncertainty level;
the AGC can balance the power imbalance left from real‐time dispatch within its power adjustment
capability range; and the proper design timescale can also avoid the frequently modulation of
dispatch system, which might exacerbate the computation burden in consequence. In other words, to
achieve the goal of power balancing, the proposed statistical functions should be utilized to properly
design the real‐time dispatch timescale.
7. Conclusions
The conclusions of this paper are as follows:
1. A negative exponential statistical function has been proposed to statistically quantify the
uncertainty level of a single IPS. Several basic properties of the statistical function have been
established.
2. To statistically quantify the uncertainty level of multiple IPSs as well as all power sources, an
SSF has been defined as the sum of multiple forecast uncertainty level functions, which can also be
represented with a single ESF with a variant time coefficient. A CSF with a constant time coefficient
has also been designed to provide a closure of the ESF.
3. Historical observations of forecast errors match well with the proposed statistical functions in
terms of temporal properties. This is not an incidental phenomenon, but reflects the general statistical
law of the prediction of future events.
4. The proposed statistical functions of forecast error have been leveraged to quantify the
statistical uncertainty levels of a single IPS, multiple IPSs, as well as all power sources, whose
parameters have been estimated based on historical data. The final CSF is a succinct representation
of the statistical uncertainty level of all power sources in the power system.
In summary, the CSF represents the overall statistical uncertainty level of all power sources. The
amplitude A and time coefficient are the two numerical characteristics that can well depict the
statistical uncertainty level.
Highlights
1. This paper establishes statistical quantification of the uncertainty levels of IPSs.
2. The proposed statistical functions have been leveraged to quantify the statistical uncertainty
levels of a single IPS, multiple IPSs, as well as all power sources.
3. Historical data sampled from real wind farms and solar power sites demonstrate the
effectiveness of the proposed functions.
References
16 of 19
1. Morris, B.; Federica, F.; Michela, L.; Dario, Z. Improvement of wind energy production through HVDC
systems. Energies 2017, 10, 157.
2. Saygin, D.; Kempener, R.; Wagner, N.; Ayuso, M.; Gielen, D. The implications for renewable energy
innovation of doubling the share of renewables in the global energy mix between 2010 and 2030. Energies
2015, 8, 5828–5865.
3. Heussen, K.; Koch, S.; Ulbig, A.; Andersson, G. Unified system‐level modeling of intermittent renewable
energy sources and energy storage for power system operation. IEEE Syst. J. 2012, 6, 140‐151.
4. Khatod, D.K.; Pant, V.; Sharma, J. Analytical approach for well‐being assessment of small autonomous
power systems with solar and wind energy sources. IEEE Trans. Energy Convers.2010, 25, 535‐545.
5. Morales, J.M.; Baringo, L.; Conejo, A.J.; Minguez, R. Probabilistic power flow with correlated wind sources.
IET Gener. Transm. Distrib.2010, 4, 641‐651.
6. Pham, H.V.; Rueda J.L.; Erlich, I. Online optimal control of reactive sources in wind power plants. IEEE
Trans. Sustain. Energy 2014, 5, 608‐616.
7. Lee, T.Y. Optimal spinning reserve for a wind‐thermal power system using EIPSO. IEEE Trans. Power Syst.
2007, 22, 1612‐1621.
8. Chen, L.; Min, Y.; Dai, Y.; Wang, M. Stability mechanism and emergency control of power system with
wind power integration. IET Renew. Power Gen. 2017, 11, 3‐9.
9. Liu, J.Y.; Zhang, L. Strategy design of hybrid energy storage system for smoothing wind power
fluctuations. Energies 2016, 9, 991.
10. Muñoz, J.; Cañizares, C.; Bhattacharya K.; Vaccaro, A. An affine arithmetic‐based method for voltage
stability assessment of power systems with intermittent generation sources. IEEE Trans. Power Syst. 2013,
28, 4475‐4487.
11. Kwon, S.; Ntaimo L.; Gautam, N. Optimal day‐ahead power procurement with renewable energy and
demand response. IEEE Trans. Power Syst. 2017, 32, 3924‐3933.
12. Ahadi, A.; Reza, S.E.; Liang, X. Probabilistic reliability evaluation for power systems with high penetration
of renewable power generation. IEEE International Conference on Industrial Technology (ICIT), Toronto,
Canada, 2017.
13. Lorca, Á.; Sun X.A. Multistage robust unit commitment with dynamic uncertainty sets and energy
storage. IEEE Trans. Power Syst. 2017, 32, 1678‐1688.
14. Hideharu, S., Tsuyoshi, F.; Nobuyuki, Y. Evaluation method for real‐time dynamic line ratings based on
line current variation model for representing forecast error of intermittent renewable generation. Energies
2017, 10, 503.
15. Guo, Y.; Zhang, B.; Wu, W.; Sun, H. Multi‐time interval power system state estimation incorporating
phasor measurements. IEEE PES General Meeting, Denver, USA, 2015.
16. Tang, Y.; Zhong, J.; Liu, J. A generation adjustment methodology considering fluctuations of loads and
renewable energy sources. IEEE Trans. Power Syst. 2016, 31, 125‐132.
17. Mai, T.; Hang, M.; Baldwin, S.; et al. Renewable electricity futures for the united states. IEEE Trans.
Sustainable Energy. 2014, 5, 372‐378.
18. Jiang, H.; Zhang, Y.; Zhang, J.J.; Gao, D.W.; Muljadi, E. Synchrophasor‐based auxiliary controller to
enhance the voltage stability of a distribution system with high renewable energy penetration. IEEE Trans.
Smart Grid 2015, 6, 2107‐2115.
19. Wang, Y.; Chiang, H.D.; Wang, T. A two‐stage method for assessment of voltage stability in power system
with renewable energy. IEEE Electrical Power & Energy Conference, Halifax, Canada, 2013.
20. Khatibzadeh, A.; Besmi, M.; Mahabadi, A.; Haghifam, M. Multi‐agent‐based controller for voltage
enhancement in AC/DC hybrid microgrid using energy storages. Energies 2017, 10, 169.
21. Li Z.; Qiu, F.; Wang, J. Data‐driven real‐time power dispatch for maximizing variable renewable
generation. Applied Energy 2016, 170, 304‐313.
22. Wei, W.; Liu F.; Mei, S. Real‐time dispatchability of bulk power systems with volatile renewable
generations. IEEE Trans. Sustainable Energy. 2015, 6, 738‐747.
23. Wei, W.; Li, N.; Wang J.; Mei, S. Estimating the probability of infeasible real‐time dispatch without exact
distributions of stochastic wind generations. IEEE Trans. Power Syst. 2016, 31, 5022‐5032.
24. Ackermann, T. Wind power in power system. John Wiley&Sons: London, UK, 2005; 365‐381.
25. Chu, Y.Q.; Li, C.C.; Wang, Y.F.; Li, J.; Li, J. Long‐term wind speed ensemble forecasting system with weather
adapted correction. Energies 2016, 9, 894.
17 of 19
26. Zhang, Z.S.; Sun, Y.Z.; Gao, D.W.; Lin, J.; Cheng, L. A versatile probability distribution model for wind
power forecast errors and its application in economic dispatch. IEEE Trans. Power Syst. 2013, 28, 3114‐3125.
27. Bessa, R.J.; Miranda V.; Gama, J. Entropy and correntropy against minimum square error in offline and
online three‐day ahead wind power forecasting. IEEE Trans. Power Syst. 2009, 24, 1657‐1666.
28. Sun, G.; Chen, Y.; Wei, Z.; et al. Day‐ahead wind speed forecasting using relevance vector machine. J. Appl.
Math. 2014, 2014, 1‐6.
29. Kavasseri, R.G.; Seetharaman, K. Day‐ahead wind speed forecasting using f ‐ARIMA models. Renew. Energy
2009, 34, 1388‐1393.
30. Xu, H.; Wu, L.; Liu, H.; Wang, R.; Cui, Z. Research on the periodicity of wind power based on the maximum
entropy spectrum estimation. International Conference on Probabilistic Methods Applied to Power
Systems (PMAPS), Beijing, China, 2016.
31. Sideratos, G.; Hatziargyriou, N.D. An advanced statistical method for wind power forecasting. IEEE Trans.
Power Syst. 2007, 22, 258‐265.
32. Jian, L.; Ying, Q.; Zong, L. Research on statistical modeling of large‐scale wind farms output fluctuations in
different special and temporal scales. Power Syst. Prot. Control 2012, 40, 7‐13.
33. Xu, X.M.; Niu, D.X.; Fu, M.; Xia, H.C.; Wu, H. A multi time scale wind power forecasting model of a chaotic
echo state network based on a hybrid algorithm of particle swarm optimization and tabu search. Energies
2015, 8, 12388–12408.
34. Mathematik, V.F. Development of a photovoltaic power prediction system for forecast horizons of several
hours. Doctoral, Carl von Ossietzky Universit¨at Oldenburg, Germany, 2016.
35. Holttinen, H.; Miettinen, J.; Sillanpää, S. Wind power forecasting accuracy and uncertainty in Finland. VTT
Technology, Espoo, 2013.
36. Bielecki, M.F. Statistical characterization of errors in wind power forecasting. Doctoral, Northern Arizona
University, USA, 2010.
37. Lauret, P.; Lorenz, E.; David, M. Solar forecasting in a challenging insular context. Atmosphere 2016, 7, 1‐17.
38. Surender, S.R.; Bijwe, P.R.; Abhyankar, A.R. Real‐time economic dispatch considering renewable power
generation variability and uncertainty over scheduling period. IEEE Syst. J., 2015, 9, 1440‐1451.
39. Silva, M.; Sousa, T.; Ramos, S.; Vale Z.; Morais, H. Distributed energy resources scheduling considering
real‐time resources forecast. 2014 IEEE PES General Meeting | Conference & Exposition, National Harbor,
MD, USA, 2014, 1‐5.
40. Li, Z.; Wu, W.; Zhang B.; Wang, B. Adjustable robust real‐time power dispatch with large‐scale wind power
integration. IEEE Trans. Sustain. Energy., 2015, 6, 357‐368.
41. He, G.; Chen, Q.; Kang, C.; Xia, Q.; Poolla, K. Cooperation of wind power and battery storage to provide
frequency regulation in power markets. IEEE Trans. Power Syst., 2016, PP, pp.1‐10.
42. Surender, S.R.; Bijwe, P.R. Day‐ahead and real time optimal power flow considering renewable energy
resources electrical power and energy systems. Electr. Power Energy Syst., 2016, 82, 400–408.
43. Wang Z.J.; Guo Z.Z. On critical timescale of real‐time power balancing in power systems with intermittent
power sources. Electrical Power System Research, 2018, 155, 246‐253.
Appendix
This Appendix contains proofs of the propositions and corollaries given in the text above.
Proposition 1. The NSFFE ( function) yields the following conservation principle
1, (i 1)
(i ) (i 1) = , (A.1)
0, (i 2)
where (i ) is the i th ‐order derivative of the function.
Proof. The first‐order derivative of the function is
1 t
(1) (t ) e , (A.2)
and the following equation holds
(1) (t ) (t ) 1 . (A.3)
18 of 19
Note that equation (A.3) holds for t 0, . Continuously differentiating the equation above
gives
(i ) (t ) (i 1) (t ) 0, (i 2) . (A.4)
This ends the proof. □
Corollary 1. The SFFE ( function) yields the following conservation principle
( i ) (t ) ( i 1) (t ) 1, (i 1)
. (A.5)
(1) (0) 0, (i 2)
Proof. Since the following equation holds for all i
(i ) t
(i ) t , (A.6)
substituting the equation above into equation (4) gives
(i ) (t ) (i 1) (t ) 1, (i 1)
. (A.7)
0, (i 2)
Note that
1
(1) ; (A.8)
(0)
therefore, equation (A.5) holds. This ends the proof. □
Proposition 2. Given n NSFFEs ( functions), the time coefficient function (t ) defined in equation (15) is
monotonically increasing.
Proof. According to equations (14) and (15), the following equation can be obtained
t
(t ) . (A.9)
ln (t )
For t1 , t2 [0, ) , if t1 t2 , then
t2 t1 C (t )
(t2 ) (t1 ) , (A.10)
ln (t2 ) ln (t1 ) ln (t1 ) ln (t2 )
where
C (t ) t2 ln (t1 ) (t1 ln (t2 )) . (A.11)
Since it is easy to prove that ln (t ) is a monotonically decreasing function, then
ln (t1 ) ln (t2 ) . (A.12)
Thus,
C (t ) 0 . (A.13)
Note that because the dominator of equation (A.10) is always positive, (t2 ) is always greater
than (t1 ) . Hence, the time coefficient function (t ) that satisfies equation (15) is monotonically
increasing. This ends the proof. □
Proposition 3. The values of the two boundary conditions of the scale factor function (t ) are
1 n
1
i
0 i 1 i , (A.14)
( ) max , ,...,
1 2 n
where
0 (0) . (A.15)
Proof. When t 0 , equation (A.9) is indefinite. Therefore, we define (0) using L’Hôpital’s rule
(t ) n
1
(0) 0 lim ( i ) 1 . (A.16)
t 0 (t )
i 1 i
Hence, the first condition of equation (A.14) holds.
When t , equation (A.9) is also indefinite. Similarly, we define ( ) using L’Hôpital’s
rule
(t )
( ) lim . (A.17)
t (t )
19 of 19
t
Note that the greater is, the slower the term e converges to zero. If
ˆ max 1 , 2 ,..., n , (A.18)
then
t
tlim (t ) lim ˆ e ˆ
t
, (A.19)
lim (t ) lim ˆ e ˆ
t
where ˆ represents the maximum among n given values of .
Thus
t
ˆ e ˆ
() lim ˆ . (A.20)
t ˆ tˆ
e
ˆ
Hence, the second condition of equation (A.14) holds. This ends the proof. □
Proposition 4. The NCSF ( function) is greater than or equal to the NESF ( function), and the maximum
deviation point t * satisfies the following equation
t* t*
1
n
e i e . (A.21)
0 i
i 1 i
Proof. The time coefficient of the NCSF ( (t , 0 ) function) is the constant 0 , the time coefficient
function (t ) of equivalent error function is monotonically increasing, and the initial value is 0 .
Therefore,
(t ) (t , 0 ) (t , (t )) 0 , (A.22)
i.e., the NCSF is not less than the NESF.
Since (0) 0 and ( ) 0 , the maximum deviation will not occur at the boundary.
Since equation (15) holds, then
t t
n
(t ) i e i
e 0
. (A.23)
i 1
i 1 i
Therefore, equation (A.21) holds. This ends the proof. □