You are on page 1of 10

Water Resour Manage (2011) 25:191–200

DOI 10.1007/s11269-010-9694-9

An Approach for Probabilistic Hydrological


Drought Forecasting

Shahab Araghinejad

Received: 7 November 2009 / Accepted: 24 June 2010 /


Published online: 17 July 2010
© Springer Science+Business Media B.V. 2010

Abstract This paper proposes an approach to monitor and forecast hydrological


drought in a probabilistic manner. The proposed approach deals with the supply and
demand variables and the role of carryover in a system to estimate the probability
of drought severity at different hydroclimatlogical conditions as well as different
storage volume levels. This approach might be of significance when the supply
and demand variables of a water resources system change considerably by climate
variation. Major probability values and their mutual use in the proposed drought
forecasting method are discussed. The presented approach is applied for the hydro-
logical drought forecasting of Zayandeh-rud river basin in Iran. This probabilistic
view of drought monitoring and forecasting is useful for risk-based decisions in water
resources planning and management. The proposed index could be used to overcome
the lack thereof in the existing surface water supply index.

Keywords Hydrological drought · Drought monitoring · Drought forecasting ·


Probabilistic methods

1 Introduction

Drought is known as the deficiency of water in one or several components of the


hydrological cycle. It occurs when the available water of a system is not sufficient
to supply at least one of the biological, economic and social water needs during a
considerable time period (Tsakiris et al. 2007). A drought plan depends on indica-
tors and triggers to characterize drought conditions and guide drought responses.

S. Araghinejad (B)
University of Tehran, Karaj, Iran
e-mail: shahab_araghinejad@yahoo.com
192 S. Araghinejad

Drought indices would be suitable if they provide reliable information to forecast


onset and termination of a drought. Despite their importance, drought forecasting
based on the combination of different indicators and triggers is usually the weak
part of drought plans, often lacking operational relevancy and scientific justification.
A problem arises when multiple indicators and triggers are specified in a drought
plan, but a systematic method for combining, using, and evaluating them is lacking
(Steinemann and Cavalcanti 2006). Indicators such as Palmer Drought Severity
Index, PDSI, (Palmer 1965), Surface Water Supply Index, SWSI, (Shafer and
Dezman 1982), and the revised SWSI (Garen 1993) have dealt with this problem
by the approach of combining various hydroclimatological variables.
The surface-water-supply index (SWSI) was introduced in Colorado in early 1980s
by Shafer and Dezman (1982) as a better indicator of water availability in the
western United States in comparison with the PDSI. Since then it has become the
most well-known index in monitoring the hydrological drought. The basic concept
of the original SWSI was the use of nonexceedance probabilities as a normalizing
technique. To improve the previous SWSI, Garen (1993) proposed a revised SWSI
as (P−50)
12
, where P = nonexceedance probability for reservoir storage + forecast
streamflow. Even in its revised form, SWSI has some shortcomings, which makes
it deficient to represent satisfactorily the hydrological drought condition of some
regions. Where water demand is a major component that controls the severity of
a drought, the original and the revised SWSI do not consider it in the classification
of drought severity. Moreover, in various systems, especially those where reservoir
storage is a significant portion of water supply, the operation of water resources
systems plays an important role in dealing with the hydrological drought. The use
of carryover storage to supply water demand in low flow periods and mitigate the
impacts of upcoming droughts, makes it important to consider the role of operating
rules in drought identification. The problem would be more significant when our
imperfect knowledge of future arises uncertainty about the factors which induce
drought, especially in the case of long-lead drought forecasting. The stochastic
behavior of drought is one of its characteristic, which is of interest in drought
monitoring and prediction (Paulo and Pereira 2007).
Recently, researches such as Rao and Voller (1997), Steinemann (2003), Draper
and Lund (2004), and Steinemann and Cavalcanti (2006) have dealt with the draw-
back of conventional drought indices using the approach of combining different indi-
cators to appropriately monitor real droughts. Rao and Voller (1997), demonstrated
that the use of multiple indices including three-month precipitation, monthly river
flow and the Palmer hydrological drought index are better indicator of drought state.
Steinemann (2003) has presented a stochastic approach for comparing, combining,
and choosing among different drought indicators using a multi-state Markov model.
Draper and Lund (2004) presented an analytical method for derivation of triggers
based on the indicators such as available water and water demand to activate optimal
hedging for reservoir operation. Steinemann and Cavalcanti (2006) used a systematic
approach to calculate drought sequences using three methods of severe drought level
method, majority of drought levels method, and the method of IN triggers and OUT
triggers.
The approach presented in this paper differs from existing methods in the way
that it suggests a systematic approach for probabilistic drought analysis based on the
An Approach for Probabilistic Hydrological Drought Forecasting 193

reliability analysis of real-time monitoring of both supply and demand components


to develop more reliable information for probabilistic drought forecasting.

1.1 Proposed Methodology

The basic concept used in the proposed approach for hydrological drought forecast-
ing is the reliability analysis by safety factor method, where Load is considered as
water demand, D, and resistance is considered as delivered water, R. Hydrological
drought occurs when 0 ≤ α < 1, where R/D = α. Obviously, small values of α
correspond to more severe droughts. Both R and D are stochastic variables which
might be presented as probabilistic distribution functions. Given the anticipated
available water (forecast streamflow plus reservoir storage), A, for a specific time
period,the occurrence probability of a drought associated with severity α, P Drα , is
calculated as
⎛ ⎞
 α D
P Drα (α) = ⎝ p (R ∩ D |A ) dR⎠dD (1)
0

where, p (R ∩ D |A ) = the joint probability function of R and D conditional toA,


and P Drα (α) = the area under the pdf, which is to the right of the R = αD line (Fig. 1).
Delivery water, R, depends on major factors such as the anticipated inflow, water
demands, reservoir storage at the beginning of the period and carryover storage.
Furthermore, water demand may have significant variation due to the same climate
variation that induces inflow as well as water loss. The joint variation of A, R, and D

Fig. 1 The occurrence


probability of a drought
of severity α
194 S. Araghinejad

leads us to calculate the occurrence probability of a drought of severity α for a future


time period by the following equation:
⎛ ⎛ ⎞ ⎞
  α D
P Drα (α) = ⎝ ⎝ p (R ∩ D ∩ A) dR⎠ dD⎠ dA
0
⎛ ⎛ ⎞ ⎞
  α D
 
= ⎝ ⎝ p(A) p (D |A ) p(R |A ∩ D) dR⎠ dD⎠ dA (2)
0

where p (R ∩ D ∩ A) = the joint probability function of forecast variables of R,


D, and A, p(A) = probability distribution function of available water (forecast
streamflow plus reservoir storage), p(D |A) = conditional probability of water
demand given anticipated available water, p (R |A ∩ D ) = conditional probability
of water delivery given the demand as well as water availability, A.
Different statistical methods could be suggested to be used through the proposed
approach. The methods proposed by Piechota et al. (2001), Araghinejad and Burn
(2005), and Araghinejad et al. (2006) could be used to estimate p(A) for a time
period in future. The p (D |A ) could be calculated by statistical analysis of historical
records of I and D in a basin, where I is the streamflow to the reservoir. It is
worth mentioning that regression methods could be useful for this analysis, when
the data of demand variability are available in a region. The calculation of demand
variability in the US is discussed in Palmer (1965). Researchers such as Vasiliadis and
Karamouz (1994) have dealt with demand driven operation of reservoirs. Moreover,
Karamouz and Araghinejad (2007) developed a model for long-term operation of
reservoirs with emphasis on drought mitigation. The methods they have proposed
could be used to develop the distribution function of p (R |A ∩ D ) for a specific time
period.

1.2 Development Procedure for the Proposed Methodology

A step-by-step procedure for monitoring and forecasting of drought using the pro-
posed approach is given herein. The variables and functions include all those
required to forecast the occurrence probability of a drought at a specific time period,
t, (say an agricultural season).
1. Identification of streamflow forecast point(s) as well as reservoir(s) in the region.
2. Implementation of the following functional relationships for the specific forecast
point and the reservoir:
◦ Available water as a function of hydroclimatological variables of the region
as well as the current storage volume of the reservoir, p(A)
◦ Demand as a function of available water at the region, p(D |A )
◦ Water delivery as a function of available water, and water demands,
p(R |D ∩ A )
3. Computation of the joint probability function of A, D, and R, using Eq. 3 at the
start of each time period, t.
p (A ∩ D ∩ R) = p(A) p(D |A ) p(R |A ∩ D ) (3)
An Approach for Probabilistic Hydrological Drought Forecasting 195

Fig. 2 The proposed algorithm for probabilistic drought monitoring and forecasting

4. Finally, computation of the area under the p(A ∩ D ∩ R) where R ≤ α D for


different values of α.

Probability distribution function of the anticipating drought is then identified as the


curve of P Drα vs. α. Figure 2 shows the algorithm for drought monitoring and
forecasting using the proposed approach.

1.3 Case Study

An example of using the proposed method is presented for the Zayandeh-rud river
basin, which is located at the central part of Iran (Fig. 3). The water resources
of Zayandeh-rud are dependent on the Zayandeh-rud reservoir. The hydrologic
situation of the Zayandeh-rud Basin is generally indicated by the inflow to the
Zayandeh-rud reservoir. The average annual inflow to this reservoir is about
1,600 million cubic meters (Mm3 ) with the standard deviation of 425 Mm3 . The
climate of the basin may be influenced by large scale climate signal of El Nino
Southern Oscillation (ENSO) as demonstrated by Araghinejad et al. (2006). ENSO
could be considered as a predictor of seasonal to annual inflow to the Zayandeh-rud
reservoir. The average annual agricultural water demand in the basin supplied by
the surface reservoir is 1,088 Mm3 with the standard deviation of 73 Mm3 . There is
also 416 Mm3 annual demands for domestic and industrial consumption of the basin.
During droughts, the restriction on water reservoir allocation is considered only for
the agricultural demands.

1.3.1 Application of the Models


As shown in the algorithm of Fig. 2, three functions are needed for the proposed
method of drought forecasting: probabilistic available water function ( p(A)); proba-
bility function of demand conditional to the occurred streamflow ( p(D |A )); and the
reservoir release as a function of demand and available water ( p(R |D ∩ A )). How
these three functions are derived for this case study is explained in the following
expressions.
An approach for making long-lead streamflow forecasts based on ENSO signal
along with the hydrological condition of the basin is applied using the geostatistical
196 S. Araghinejad

Fig. 3 The map of Zayandeh-rud river basin

procedure of forecasting (GBPF) developed by Araghinejad et al. (2006). GBPF


quantifies forecast uncertainty through the estimation variance and is capable of
producing a continuous normal probability distribution function of the predicted
value considering multiple predictors. In the case study, it is understood by the sta-
tistical analysis that at least a 3-month period (October to December) of observation
is needed for an appropriate estimation of annual drought situation of the basin.
Southern Oscillation Index (SOI), a well known index of ENSO, as well as observed
inflow and rainfall from October to December are used as predictors of the sum of
9-month (January to September) inflow to the reservoir.
The agricultural demand variable of the Zayandeh-rud basin is negatively cor-
related to the inflow to the reservoir. It could be demonstrated by the scatter plot
of the annual inflow to the reservoir versus the annual agricultural demand of the
Zayandeh-rud basin for ten years of 1990 to 2000 as shown in Fig. 4. As GBPF
produces forecast variables as a normal distribution, the conditional probability of
demand variable ( p(D |I) ) for a specific streamflow value (i) is also forecasted as
a normal distribution with the average of μ D|I and the standard deviation of σ D|I ,
which are calculated by the following equations, respectively:

σ D (i − μ I )
μ D|I = μ D + ×ρ (4)
σI

σ D|I = 1 − ρ2 σD
2
(5)
An Approach for Probabilistic Hydrological Drought Forecasting 197

Fig. 4 Annual inflow to the reservoir versus annual agricultural demand in the case study

where, ρ = correlation coefficient between inflow to the reservoir and the agri-
cultural demand; μ D and σ D = average and standard deviation of agricultural
demand, respectively; μ I and σ I = average and standard deviation of forecast inflow
to the reservoir, respectively; and i is the specific value of the forecasted inflow
to the reservoir. More details on joint probability distributions could be found in
Yue (2000). p(D |I) could be considered as the probability occurrence of demand
conditional to the available water, p(D |A) , where the deterministic variable of
current storage volume is added to the streamflow forecast.
The optimum release from a surface reservoir could be obtained by different
operating rules. The rule considered for the Zayandeh-rud reservoir in this study
is the hedging rule developed by Karamouz and Araghinejad (2007), which activates
restrictions on allocating water to agricultural demands during droughts. Using that
operating rule, the optimum release (R) is calculated based on the specific values of
available water.

2 Results and Discussion

As an illustrative example, the procedure of drought forecasting in year 1997 is


presented here. As it has mentioned earlier, observed predictors of October to
December are used to forecast variables in the period of January to September. The
average 9-month inflow of January to September is 1,425 Mm3 . In 1997 the 9-month
inflow to the reservoir is estimated by the GBPF model as a normal distribution
198

Table 1 An example of drought forecasting using the proposed approach


Probable forecast 961 to 1,004 1,004 to 1,947 1,047 to 1,090 1,090 to 1,133 1,133 to 1,176 1,176 to 1,219 Expected
inflow (Mm3 ) probability
Category indicator 982 = 1,025.5 = 1,068.5 = 1,111.5 = 1,154.5 = 1,197.5 = of the
(Mm3 ) (961 + 1,004)/2 (1,004 + 1,047)/2 (1,047 + 1,090)/2 (1,133 + 1,090)/2 (1,176 + 1,133)/2 (1,219 + 1,176)/2 proposed
Exceedance 0.0215 0.1359 0.3413 0.3413 0.1359 0.021 index
probability
Average demand 1,231.09 1,182.5 1,134.8 1,086.9 1,039.1 991.3
(Mm3 )
Standard deviation 72.58 72.58 72.58 72.58 72.58 72.58
of demand (Mm3 )
Water delivery (R) 927 974.5 1,015.5 1,052.5 1,104.5 1,158.5
(Mm3 )
α ≥ 1 = p(R>D) 0 0.01 0.09 0.3192 0.9 0.99 0.28
α < 1 = p(R<D) 1 0.99 0.91 0.6808 0.1 0.01 0.72
S. Araghinejad
An Approach for Probabilistic Hydrological Drought Forecasting 199

Table 2 Results of probabilistic drought forecasting in the case study


Year Actual Expected Proposed index
streamflow value of the Probability of Probability of Probability of Probability of
forecast α≥ 1 0.8 ≤ α< 1 0.5 ≤ α < 0.8 α≤ 0.5
streamflow
1997 944 1,090.0 0.28 0.53 0.17 0.02
1998 1,340 1,250.0 0.48 0.37 0.14 0.01
1999 720 810 0.21 0.37 0.34 0.08
2000 680 730 0.04 0.18 0.66 0.12

with the average of 1,090 Mm3 and the standard deviation of 43 Mm3 . The storage
volume at the start of January is considered as 420 Mm3 . Table 1 shows the
steps of computing probability values as well as the α index. For the simplicity,
p(I) is considered as six categories within the range of inflow confidence interval
considering an indicator for each category. The delivery water, R, is also calculated
in a deterministic manner. For each streamflow and demand, the optimum release is
obtained and compared with the demand. For instance, for the streamflow interval
of 1,047 to 1,090 Mm3 the expected demand is obtained as a normal distribution
with an average of 1,134.8 and standard deviation of 72.58 Mm3 associated to the
indicator value of 1,068.5 Mm3 inflow to the reservoir. The optimum associated
release obtained by the operating policy is calculated as 1,015.5 Mm3 . The proposed
drought index is then calculates as the probability of drought occurrence (α < 1)
as 0.91 and the probability of non-occurrence of drought (α ≥ 1) as 0.09. It should be
notified that the specific probability of this category is about 0.34, which should be
considered in the calculation of the final expected value of the drought occurrence.
This method is thoroughly performed for the range of forecast inflow and demand in-
tervals by an algorithmic procedure. Finally, for the case of year 1997 the probability
of occurring drought is obtained as 0.72.
The results of drought forecasting in the Zayandeh-rud basin from 1997 to 2002
are shown in Table 2 for different drought severities. As it is demonstrated by Table 2
the most severe drought is forecasted for the case of year 2000 and the less severe
drought is forecasted in year 1998. The results are in a good agreement with the
drought loss reported for the region; however, the major benefit of the proposed
method is to make risk-based analysis possible for the system.

3 Conclusion

An approach has been proposed for drought monitoring and forecasting in the
regions, where reservoir storage plays an important role in water supply and drought
mitigation and water demands vary significantly due to the climate variation. The
proposed approach uses the reliability analysis of water demand and water delivery
in a probabilistic manner to estimate the probability values of a drought at a specific
time period. The aim of this paper has been to define an approach as meaningfully
as possible to overcome the lack thereof in the existing supply water severity index.
Hydrological forecasting model, reservoir operation model, demand variability func-
tion as well as the storage capacity of a system are major components of the drought
forecasting procedure of this paper. Examples of applying the proposed method
200 S. Araghinejad

were presented for the Zayandeh-rud basin in Iran. The method is suitable to be
used through a risk-based decision making process helping decide on restriction on
demand allocation as well as target storage volume at the end of a decision period.
The proposed method has potential utility for conditioning a water resources related
outlook, in the drought state of the basin.

Acknowledgements Two anonymous reviewers are thanked for their comments that have en-
hanced this paper.

References

Araghinejad S, Burn DH (2005) Probabilistic forecasting of hydrological variables using geostatistics.


Hydrol Sci J 50(3):837–856
Araghinejad S, Burn DH, Karamouz M (2006) Long-lead probabilistic forecasting of streamflow
using ocean-atmospheric and hydrological predictors. Water Resour Res 42(3):W03431, 1–11.
doi:10.1029/2004WR003853
Draper AJ, Lund JR (2004) Optimal hedging and carryover storage value. J Water Resour Plan
Manage 130(1):84–87
Garen DC (1993) Revised surface-water supply index for western United States. J Water Resour
Plan Manage 119(4):437–454
Karamouz M, Araghinejad S (2007) Drought mitigation through long-term operation of reservoirs:
a case study. ASCE J Irrig Drain 134(4):471–478
Palmer WC (1965) Meteorological drought. Research paper no 45. US Department of Commerce
Weather Bureau, Washington
Paulo AA, Pereira LS (2007) Prediction of SPI drought class transitions using markov chains. Water
Resour Manag 21:1813–1827. doi:10.1007/s11269-006-9129-9
Piechota TC, Chiew FHS, Dracup JA, McMahon TA (2001) Development of exceedence probability
streamflow forecast. J Hydrol Eng 6(1):20–28
Rao AR, Voller TL (1997) Development and testing of drought indicators. Water Resour Manag
11:119–136
Shafer BA, Dezman LE (1982) Development of a Surface Water Supply Index (SWSI) to assess
the severity of drought conditions in snowpack runoff areas. In: Proc western snow conference,
Colorado State University, Fort Collins, CO, pp 164–175
Steinemann A (2003) Drought indicators and triggers: a stochastic approach to evaluation. J Am
Water Resour Assoc 39(5):1217–1234
Steinemann A, Cavalcanti L (2006) Developing multiple indicators and triggers for drought plans.
J Water Resour Plan Manage 132(3):164–174
Tsakiris G, Pangalou D, Vangelis H (2007) Regional drought assessment based on the reconnais-
sance drought index (RDI). Water Resour Manag 21:821–833. doi:10.1007/s11269-006-9105-4
Vasiliadis HV, Karamouz M (1994) Demand-driven operation of reservoirs using uncertainty-based
optimal operation policies. J Water Resour Plan Manage Div Am Soc Civ Eng 120(1):101–114
Yue S (2000) Joint probability distribution of annual maximum storm peaks and amounts as repre-
sented by daily rainfalls. Hydrol Sci J 45(2):315–326

You might also like