0 évaluation0% ont trouvé ce document utile (0 vote)

7 vues22 pagesnawaz and Adeloye Rainfall and Runoff model

79e41510d7de0a1c1e

© Attribution Non-Commercial (BY-NC)

PDF, TXT ou lisez en ligne sur Scribd

nawaz and Adeloye Rainfall and Runoff model

Attribution Non-Commercial (BY-NC)

0 évaluation0% ont trouvé ce document utile (0 vote)

7 vues22 pages79e41510d7de0a1c1e

nawaz and Adeloye Rainfall and Runoff model

Attribution Non-Commercial (BY-NC)

Vous êtes sur la page 1sur 22

113

Evaluation of monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

Department of Civil and Offshore Engineering, Heriot-Watt University, Edinburgh EH14 4AS, UK e-mail: a.j.adeloye@hw.ac.uk Abstract Results of further evaluation tests of a monthly rainfall-runoff model used for extending streamflow data records in England and Wales are presented. In particular, since the objective of the record extension exercise was to make available long enough data records for reservoir yield assessment, model performance in reproducing the reservoir storage-yield-reliability relationship during calibration was examined. Finally, it can be argued that this type of test should be among the suite of performance checks on rainfall-runoff models in situations where low flow is the primary concern.

Estimation de l'coulement mensuel par un modle rgressif pluiedbit utilis pour l'valuation du dbit fourni par un rservoir

Rsum Cet article prsente les rsultats d'une procdure d'valuation d'un modle rgressif pluie-dbit l'chelle mensuelle utilis pour tendre des sries chronologiques de dbit en Angleterre et au Pays de Galles. Comme l'objectif de l'extension de donnes tait de produire des sries suffisamment longues pour valuer le dbit fourni par un rservoir, nous avons tudi la capacit du modle reproduire la relation entre stockage, lachure et fiabilit dans sa phase de calage. Nous pensons finalement que ce type de test devrait faire partie de la batterie d'essais auxquels devraient tre soumis les modles pluie-dbit lorsque les dbits d'tiage constituent la principale proccupation.

INTRODUCTION It is seldom the case that adequate historic streamflow data are available for the planning of reservoirs and so one of the main tasks during such activities is to extend any available short record prior to the reservoir planning analysis. Planning is used here to mean the determination of storage capacity required at the reservoir site to meet a given demand with a specified level of reliability (Adeloye, 1996). For such exercises, long records which have the potential of including very severe drought sequences are often preferred because the uncertainty associated with the derived storage-yield-reliability relationship can be small. The Environment Agency (EA) in England and Wales recently recommended the use of a Behaviour Analysis approach for estimating the yield of reservoir systems from historic streamflow data records (Drayton & Lambert, 1995). Carty & Cunnane (1990) investigated the performance of various empirical techniques for storage-yield analysis and concluded that a behaviour analysis was superior to most others because

Open for discussion until 1 August 1999

114

it resulted in the least bias and standard error for reservoir capacity estimates. However, to ensure that the yield analysis incorporates the worst possible historic drought sequence and hence to improve the future dependability of yield estimates, the EA also recommended that analysis should be based on a long trace of historical record, specifically spanning the period 1920 to date. Only a small number of catchments in Britain have streamfiow records that are longer than half this length, implying that extensions of the available short records will have to be carried out for the methodology to be applicable. There are several methods for extending streamfiow data records, ranging from simple rainfall-runoff correlation, to the complex analysis involving the physics of catchment dynamics and its response to rainfall inputs (Hirsch, 1979, 1982; Jamieson, 1994; Raman et al., 1995; Xu & Singh, 1998). The particular model whose reconstructed streamfiow data were evaluated here was developed by Wright (1978) and has also been recommended by the EA for record extension for the purpose of yield assessment in England and Wales. The model is based on an extension of rainfall-runoff correlation techniques and is particularly convenient because for most catchments in England and Wales, there are considerably longer rainfall records, often much earlier than 1920, which could form the basis for extending short streamfiow data records. However, before using such a model for record extension, it is important to ensure that the model performs adequately during calibration and verification, i.e. its reconstructed data record must reproduce adequately the relevant characteristics of the historic record. Customarily, model adequacy is assessed by comparing various streamfiow moments, such as the annual and monthly mean, standard deviation, skewness, serial correlation, of the reconstructed and observed data records over a concurrent calibration period. This task has been largely accomplished with satisfactory results in several applications of Wright's model (Jones, 1983; Jones & Lister, 1995). However, since the objective of the EA in reconstructing these data is to use them for reservoir storage-yield-reliability analysis, an alternative, superior approach would be to compare the storage-yield-reliability relationship of the observed and reconstructed data records, particularly during the calibration period as suggested by several workers (e.g. Hirsch, 1979; Raman etal., 1995; Savic etal., 1989). Adeloye (1990) noted that, while the mean, standard deviation and other commonly used moments of streamfiow can affect reservoir storage, the relationship between all these variables and storage is often complex so that one cannot easily predict the behaviour of reservoir storage by merely observing, in isolation, any of the relevant basic streamfiow statistics. The reservoir storage-yield-reliability relationship integrates the combined effects of all the relevant streamfiow statistics and hence represents the best criterion that should be used for model verification in such a circumstance. This paper describes the results of such an adequacy test on Wright's model. Historic streamfiow data records and their concurrent reconstructed records for eleven catchments in England and Wales were obtained from the Climatic Research Unit, University of East Anglia, UK. (P. D. Jones, personal communication) and the EA. The sizes of the catchments vary from 400 to 10 000 km2 and they are located in both lowland and upland Britain. For the study, reservoir storage-yield-reliability

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

\\5

relationships were derived for each of the records using a simulation approach (McMahon & Mein, 1986). By comparing the storages obtained from the reconstructed and observed records over the concurrent calibration period, it was possible to arrive at errors in storage estimates due to the calibrated model flow reconstructions.

WRIGHT'S MONTHLY RAINFALL-RUNOFF REGRESSION MODEL A detailed description of the model by Wright is available elsewhere (Jones, 1983, 1984; Wright, 1978); only its main elements are presented here. The general form of the monthly model is (Wright, 1978): log10F = C+ >,.*,. + fjblGl

/=-/, j=-i2

+ j^c.M,

*=-/,

+ dEB + eESMD

(1)

where F0 is the flow for the current month (m3 s"1); C is the intercept (constant term of the regression); R4 are the residual rainfall values (precipitation-actual vapotranspiration, including the negative sign) for the current and preceding /, months (mm); G_j are the modified residual rainfall values (see the next paragraph for its derivation) for the current and preceding Z2 months; and M.k are the soil moisture deficits at the end of the current and preceding Z 3 months (mm). The term EB is a flood factor for high values of modified residual rainfall (Ga) for the current month, to improve estimates of high flows; and ESMD is a low flow factor for high values of the soil moisture deficit (M) for the current month, to improve estimates of low flows. The modified residual rainfall (G) terms allow for the effect of antecedent soil moisture content on the runoff and were determined in the model using the following rules: G_I = (R/300)[(BR_I/100)+A+M_,_1+R_i_i] G_f = 0 for R_, >0 for R_: < 0

where A and B are coefficients determined by error mapping techniques as will be described in a later section. The term EB is assumed equal to G0, the modified residual rainfall for the current month, only when G0 exceeds some arbitrarily set threshold EB0; otherwise EB is zero. Similarly, ESMD is assumed equal to M0 when MC) is less than ESMD0, a pre-defined threshold value; otherwise it is zero. The R4 variables are also constrained to a maximum value (RMAX), depending on the value of the lag i as follows: if / > MAXLAG, then R_t is set to RMAX if its value exceeds RMAX. This is done to ensure that residual rainfall greater than RMAX occurring in the distant past does not contribute to flow in the current month. Values of EB0, ESMD0, MAXLAG and RMAX are, like coefficients A and B, obtained by error mapping. The error mapping involves a trial-and-error process of assuming values for these constants, running the model and noting the standard error of estimate. Different trial values are used and an error contour surface is generated; the

116

100

200 A Constant

300

100c o o

50235

150 100 ESMD Constant Fig. 1 Error mapping for six constants: Thames to Eynsham (Wright, 1978).

50

combination of constants giving the minimum error is then taken as the solution. Wright (1978) recommends optimizing two related constants together in one analysis, beginning with RMAX and MAXLAG; A and B; and finally EB0 and ESMD,,, Figure 1, taken from Wright (1978), is an example of an error map generated for the

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment \ \ 7

Thames at Eynsham catchment. The optimized values for RMAX and MAXLAG were found to be 5 mm and 12 respectively for a model standard error of estimate equal to 22, implying that values of residual rainfall greater than 5 mm have no significant effect upon river flows recorded more than 12 months after the rain event. However, as shown in Fig. 1(a), increasing RMAX to 100 mm will only increase the model error by about 2%, which means that the model results are not very sensitive to this parameter and it may effectively be omitted from the model. Of course, once RMAX is dropped then MAXLAG becomes irrelevant. Similarly, in Fig. 1(b) it is apparent that the constants A and B could vary widely, in this case between 160-300 and 50100 respectively, without affecting the model error. In the same vein, Fig. 1(c) shows that EB is in fact not a significant parameter for the Thames at Eynsham catchment; similar observation was recorded for most of the catchments analysed in Wright (1978). Thus although the model of Wright (1978) has many parameters, its performance is relatively insensitive to large variations in these parameters; advantage can be taken of this feature to omit the parameters from the model and, by so doing, simplify both its structure and calibration. This was indeed the case for most of the catchments analysed in Wright (1978). Model identification will produce the number of lags (/,, l2, and /3) and the coefficients ah bp ck, (and also d and e but only when EB and ESMD are included in the regression model) (Wright, 1978). For most of the catchments that have been investigated so far, the general model has been found to simplify to an equation incorporating between five and 15 terms, the precise number depending on the particular catchment (Jones, 1983; Jones & Lister, 1995). For example, certain catchments in the south of England tend to have a delayed runoff response, thus requiring a greater number of terms to accommodate the dominant influence of groundwater contribution to river discharge in such catchments. The primary inputs to the model are the monthly flow (m3 s '), the monthly catchment rainfall (mm), and monthly actual vapotranspiration data (mm); the latter two are used to derive the residual rainfall, which is the main independent variable in the multiple regression formulation. Another independent variable, i.e. the soil moisture deficit (SMD), is also not input directly but derived from the residual rainfall. The variable SMD is set up whenever the residual rainfall is negative, implying that some of the actual vapotranspiration is being taken out from the available soil moisture capacity, resulting in the lowering of the soil moisture content level below the field capacity. Thus the derived values of the SMD are a mixture of zeros and numbers, depicting respectively periods of wetness and dryness. In Jones & Lister (1995), rainfall data were obtained from the UK Meteorological Office (UKMO) archives, while the flow data were obtained from the appropriate regions of the EA. As for actual vapotranspiration data, Jones & Lister (1995) used the seasonally (i.e. monthly) constant values recommended in Wright (1978), which had been obtained through a trial-and-error optimization using a catchment model (see Wright, 1978: Appendix H4, p. 67). While it is theoretically possible to derive both the potential and actual vapotranspiration by scaling from Penman open water surface evaporation estimates, Wright (1978) found that using the actual vapotranspiration obtained from such an approach resulted in poor flow reconstructions. This was

118

attributed to the significant seasonal differences he observed between the actual vapotranspiration derived from the Penman approach and that to be expected from a catchment water balance study. Other advantages of using seasonally constant actual vapotranspiration values are its simplicity and removal of the extra effort involved in collecting the meteorological data necessary for evaluating the Penman expression.

MODEL EVALUATION The catchments The adequacy tests were based on eleven catchments in England and Wales as shown in Fig. 2. The concurrent observed and reconstructed data at these sites were compiled from Jones & Lister (1995). Table 1 contains a summary of the catchments and their data records.

1 : Eden to Temple Sowerby (catchment of 2); 2: Eden to Warwick Bridge; 3: Exe to Thorverton; 4 : Great Ouse to Denver Complex; 5: Thames to Eynsham (also close Teddington); 6: Tees to Broken Scar; 7: Wensum to Costessey Mill; 8: Wharfe to Addingham; 9: Wye to Redbrook; 10: Tyne to Bywell.

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

\ \9

In general, observed streamflow records for the catchments are available for the period 1970-1993. The only exception was the Thames to Teddington catchment where the observed data record was much longer, from 1885 to 1976. These periods include the exceptionally dry event of 1975/1976 in England and Wales, which is significant for yield analysis based on a critical period approach. It is on record that, during this year, most of the catchments in northern England and parts of East Anglia and the southwest, which together contain eight of the 11 catchments in this investigation, received only about 70% of the long-term average rainfall. Parts of southern England and the Midlands where the Thames and Ouse catchments are located received even lower rainfall during 1975/1976, just below 50% of the average (Wright, 1978). Jones & Lister (1995) have reported the results of the calibration and verification of Wright's model for each of these catchments, including the use of the verified models to extend historic records back in time. However, these previous verification exercises did not consider the storage-yieldreliability characteristics of the records; hence the need for the current study. The periods over which the data records have been reconstructed for each catchment are also shown in Table 1, although our evaluation of the model performance was only based on the calibration period, in line with earlier evaluation of the model by previous workers. The variability of the flows, as characterized by the coefficient of variation (CV) of annual flows, is shown in Table 1. The CV is the most important parameter of the streamflow process influencing reservoir capacity (Adeloye, 1990; Burges & Linsley, 1971). In the absence of nonstationary effects caused by climate and landuse changes, an assumption which is being made herein, flow variability is primarily a function of catchment geology and rainfall characteristics. For England and Wales, catchment geology is the dominant factor (Wright, 1978). Catchments having relatively pervious geology are able to sustain reasonable base flows in their rivers, thus helping to reduce flow variability. Significant rainfall variability will counter this. Conversely, the variability of flows in rivers underlain by an impervious geology is totally a function of rainfall variability. Examples of these two extreme conditions are available in the eleven catchments, with the catchment of River Eden belonging to the former and that of Thames (Eynsham) belonging to the latter. The value of the annual CV also gives an indication of the relative significance of within-year and over-year storage requirements and hence the resilience of a reservoir system. Vogel & Bolognese (1995) characterized the resilience of surface water reservoir systems by a parameter m, which is related to the annual CV and the yield as follows:

1 a

m=

n\ (Z)

CV where a is the yield as fraction of the mean annual flow. According to Vogel & Bolognese (1995), systems dominated by over-year storage requirements have values of m below 1 ; such systems exhibit a high CV, a high a or both, and are not resilient because they will take several years to recover following any significant drawdown. Within-year systems are characterized by a value of m over unity; the higher this

120

&

O i ^ t J O O r t O O H l M h ^ r m O N o o - H O i r ) i > o o o

O-

s if

Oit* S

0 0 - * O N N O r ~ T j - o t - ~ I ~ -

l o O M O m n t n - i m

oS

c")<mONOoNNOint~~No*-cn

as 5

lO ^H ^H NO o o o N O O i o O o o o N o m O N l^ i o o r n v - > O N r - ( N o O ' a - - ^ m ^ > ^Tj-t-~-oON\OTt"OOOOc>->c""> M -H -H M

oo oo * i n N O O N r ^ N O o t ^ N O O r ^ ^ O C N I f N o o c ^ ^ c n ^ n o o N o o O N O N O c ^ c n N O O N O N o o i n t o m ' ^ - c n

S<2< >_

CN

*-<

*-"<

C N

o o o o o o o o o o o

O i ^ N f o ^ E ^ O \ O o o N 0 \ o O t ^ r O ( N r ^ r ^ N O c * ~ i s

X> n

<U

lH

t/1

o

** ->

(fl

ord

<%

a o o

T3

ro J3

O

T-1

O N O N O N O N O N O N O N O N O N O N O O

N Q^ o -^

< u

flow truc

0) ^ T3 n U U

c ^ m m c ^ m r * ~ i c ^ i m m m ^ o ONONONONONONONONONONt^ ONONONONONONONONONONON O N O m O N i o ^ - o o N o o o ^ - i r > N o m N O N O n T f i n ' o m ' n Q O o o o o o o i ^ o o o o o o o o o o o o o o

a

-S

c

f>

O .3 X l J u ' .3 3 d-c S

ON J O

oo Pi

eu

O O O ^ l - t O ' O ^ * m

O O 0 0 N 0 ^ - * N 0 C " ) O ~ o o t N t - ~ r t r t O c n - ! t t ^ N O c n o o N O N O s r i T f O O (N - H ^ ON -O _

T ^ S S-3

" 3

K

c ^ S

O O

o

X> Pi

O

am ton

SP c S3 " S "O ^3 < H

r-> TO

rton

1

U

S o

u --- w < D 3 es 1 3 <>

-S h o 5 P

s a>

m H S

> , u C* <G

<"

y o

fi m o) > . J 3 "O

u -a

x S s -a w ^ t? H

OJ

-3

>

O H H W H W

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

\ 21

value, the higher is the resilience or the ability of the system to bounce back following a failure. The significance of this kind of preliminary insight into the likely behaviour of a reservoir system is that it gives a pointer as to the level of detail of the data required for storage-yield analysis. While purely over-year storage systems, i.e. those exhibiting m values close to zero, may be adequately analysed using annual time series flow data, the analysis effort required for purely within-year systems, i.e. m I, can also be significantly reduced by identifying and analysing only the "critical 12 months" of the data record. However, since most systems exhibit a mixture of both features, albeit to varying degrees, analysis of the complete monthly time series data is often required. This is likely to be the case for the catchments in Table 1 given their low-to-moderate annual CV values.

Reservoir storage-yield-reliability analysis Reservoir storage-yield-reliability analysis was by simulation, based on the following reservoir mass balance equation: St+l = S, + Q,-Dl-E,-Ll; 0<Sl+l<K (3) where 5,+, and S, are the volumetric storages at the beginning and end of period t respectively; Q, is the volumetric inflow during t; D, is the actual volumetric draft during t; E, is the volumetric net evaporation loss during t; L, are the volumes of other losses during t; and Ka is the active storage capacity. For the main analysis, the effect of evaporation loss was ignored; however, subsequent sensitivity tests examined the impacts of evaporation. When it has been considered, volumetric evaporation loss (equation (4b)) was introduced using a linear approximation to the reservoir surface area-storage relationship (equation (4a); see also Loucks et al., 1981; Gan & McMahon, 1991): A, = aS, + b E, = en,(A, + Al+l)/2 (4a) (4b)

where A, and Al+i are the surface areas at the beginning and end of t respectively; en, is the open water surface evaporation (mm) during t; and a and b are constants. The other losses L such as seepage and conveyance losses, were totally ignored. Incorporating equations (4a) and (4b) into equation (3) will yield: Sl+I = , * JS, (l - 0.5a(e, )) + & - > , - bien, )] J \l + 0.5a(enl )] L (5)

Given time series of flow data (Q t = 1, 2, ..., T, where T is the total number of data points) at a hypothetical or real reservoir site and the corresponding open water surface evaporation, the determination of Ka for known yield and reliability using equation (3) or equation (5) is a trial-and-error approach. First, a trial value of Ka is assumed and the appropriate equation is solved recursively for an initially full reservoir until period T is reached. (The assumption of an initially full reservoir was

122

tested later in the sensitivity analysis.) During the recursion, note is taken of any period t for which the reservoir failed, i.e. when the reservoir was unable to supply the full yield and hence ended up being empty, and the volume of the shortfall in that period. Then system performance or reliability is determined and if this is different from that desired, a new value of Ka is assumed and the procedure repeated until the solution converges to the desired reliability. The relatively short flow records (typically 24 years) to be used in the behaviour analysis are such that the capacity estimates may deviate from their stationary mean value (Pretto et al., 1997). However, for the present objective, whether or not the capacity becomes stationary is not important since the concern will be with the difference in capacity estimates for records of the same length: any influence of the record length will be reflected equally in both estimates and should therefore cancel out when differences are taken. For the reservoir planning, three types of system reliability as proposed by Klemes et al. (1981) were considered: (a) Annual (or occurrence-based) reliability: " number of non-failure years number of years in record

(b) Time-based reliability: duration of non-failure periods ' total duration of record, T

(c) Volumetric (or quantity-based) reliability: '' actual volume of water supplied total volume demanded over T

In general, where the three measures of reliability are obtained for any Ka, then Rv > R, > Ra. Finally, once capacity is obtained for the concurrent observed and reconstructed data, the relative error due to the reconstructed data is then estimated using: E(%) = lOOKar~Kaa (7)

where E is the relative error (%), Km is the observed storage capacity and Kar is the corresponding storage capacity obtained from the concurrent reconstructed data record.

RESULTS Reservoir storage-yield-reliability curves The derived storage-yield-reliability relationships obtained for the smaller ten catchments are shown in Fig. 3. As the yield increases, the storage requirement also

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment 123

EXE (THORVERTON)

OUSE (DENVER)

THAMES (EYNSHAM)

. . -A.-e

90

100

WHARFE (ADDINGHAM)

.

J

A-J^J^^.

A>

II II II

^-A__4

1111

. . I I

Storage (% MAF)

124

increases, first gradually before flattening out as the yield approaches the mean flow. In this flat region, a very large storage increase is required to meet a modest increase in yield. This observation was also made by Adeloye (1996), who noted that any uncertainty which results in the storage being over-designed in the flat region does not yield any significant advantage in terms of additional yield; instead, such an error will be accompanied by huge additional reservoir storage spaces. The shape of the storageyield curves also gives an indication of the relative costs of reservoir development at the various sites. For example, the more steep the storage-yield relationship, the lower is the reservoir storage (as fraction of mean flow) required to meet a given level of demand and hence the lower the reservoir construction cost (assuming material and labour costs do not vary widely between the various sites). Furthermore, the less the reservoir storage capacity, the lower will be the exposed surface area and any evaporation fluxes through this surface. The impacts of direct evaporation fluxes will be discussed in a later section dealing with the sensitivity studies. Another point which has become even more evident in the storage-yieldreliability functions is the relationship between the two reliability measures Rv and R,. In general, the curve for Rv of 98% always lies everywhere above that for R, of 98%. This would imply that for any yield level, the reservoir storage capacity for Rv is lower than that for a similar value of Rr Put differently, a reservoir capacity designed for Rt of 98% would give a higher volumetric reliability or alternatively would supply a higher yield at a volumetric reliability of 98 %. This is an interesting point, particularly when it is realized that current levels of service for raw water availability in England are specified in terms of time-based reliability. The problem with time-based reliability is that it does not distinguish between a failure which only produced a shortfall of 1 % and that resulting in a shortfall of 99% : both are regarded as failures. The volumetric reliability, on the other hand, takes into account the actual volume of shortfall and is therefore a more appropriate measure to use when the desire is to assess the real impact of water shortage.

Model adequacy: use of storage-yield relationships vs other criteria The relative errors computed according to equation (7) are plotted in Fig. 4 for all the 10 smaller catchments. Similar information for the larger Thames to Teddington catchment is shown in Fig. 5. Where the relative error is positive, i.e. when the reconstructed data have over-predicted the storage of the observed data, this has been designated "oe" in Fig. 4(a)-(c); otherwise it is an underestimate. Another notation appearing in the figure is "n/a", i.e. not available, which indicates a situation where 98% time-based reliability is not feasible for the given yield. This was the case only for the low yield of 30% of the mean flow, where the time-based reliability for zero storage was almost 100% in some cases. For the three yield levels of 0.3, 0.6 and 0.9 of mean flow, there were only two river catchments (River Eden at Warwick Bridge and River Exe at Thorverton) for which the absolute value of the relative error was 10% or below. Srikanthan & McMahon(1980) suggested an error limit of 10% as being acceptable for an exercise

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment 125

(a)

(b)

(c)

Catchment no. 1: Eden (Temple Sowerby); 2: Eden (Warwick Bridge); 3: Exe (Thorverton); 4: Ouse (Denver Complex); 5: Thames (Eynsham); 6: Tees (Broken Scar); 7: Wensum (Costessey Mill); 8: Wharfe (Addingham); 9: Wye (Redbrook); 10:Tyne (Bywell). n/a data not available; oe: over-estimation of storage capacity, otherwise an under-estimate,

Fig. 4 Relative errors for 10 catchments during calibration (1970-1993)- (a) 30% yield, (b) 60% yield, and (c) 90% yield.

of this nature and, on this basis, only the calibrated and verified models for these two catchments could be considered adequate for reconstructing flow data for reservoir storage-yield analysis. The performance of the model at the other sites is very poor. For example, the results for catchment 6 (River Tees at Broken Scar) clearly demonstrate the scale of the problem if a data reconstruction model is not adequately reproducing the droughts of the observed data, which is why the storage-yield functions are being poorly reproduced. For this catchment, for a yield of 90% of the mean flow, the underestimate in storage is either 68 or 69% depending on whether it is based on Rv or R, respectively. For a much smaller yield of 30% of mean flow, the error is an overestimate of 140% for an R value of 98%.

126

100

< 5

0s-

80 60

Jit

- A ^ B L^^^

_,

-Rv=98%(obs) -Rb=98%(obs) - - A- - -Rv=98%(rec) - - O - Rb=98% (rec)

Reliability = 98%

~"~ 40

2

CD

>

1 20 I I 0 II 0 100 i

1-

.

200 250 Reliability = 95%

50

100

150

- Rv=95% (obs) Rb=95%(obs) - A- - - Rv=95% (rec) - O - - Rb=95% (rec) 20 100 Reliability = 90% 40 60 80 100

Storage (% MAF)

03 tu -40

IT

60

90

60

90

YIELD (% AMF)

Fig. 5 Storage-yield-reliability relationships and relative errors for Thames to Teddington catchment during calibration (1886-1976).

The high relative errors are indicative of the deficiency of Wright's model in reconstructing streamflow data for yield analysis, a deficiency which is not obvious when the performance of the model is evaluated only in terms of basic streamflow statistics. For example, the relative errors in the mean flow for the same eleven catchments have been evaluated from the results in Jones & Lister (1995) and are presented in Table 1. As far as the mean flow is concerned, Wright's model has performed excellently in all the catchments except one, the Great Ouse at Denver

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

\ 27

Complex (catchment 4), where the relative error is above 10%; however, even for this catchment the relative error of 16.4% in the mean flow is not as high as those associated with the storage-yield relationship. Wright (1978), in his pioneering work on the catchment rainfall-runoff model, also examined basic statistics such as the mean, variance and skew coefficient, and concluded that these were well preserved. Indeed, the relative errors for the mean flow in Wright (1978) were broadly similar to those of Jones & Lister (1995). Jones & Lister (1995) also compared the 1-month cumulative low flow which is exceeded 90% of the time, i.e. Q90; the estimated relative errors associated with these are also included in Table 1. The relative errors associated with Q90 presented in Table 1 are much higher than those of the mean flow, which is not surprising given that ?90 is much more related to storage requirements than the mean flow. However, restricting such an analysis to just one month assumes that the critical period of any reservoir to be built is one month which is unlikely to be the case. Wright (1978) did examine durations of up to eighteen months; however, the results were not presented in such a way as to allow the relative errors for the Q9Q of the higher durations to be readily estimated. Furthermore, flow analysis to determine low flow quantiles such as the Q90 can introduce other uncertainties, particularly the model error associated with the choice of the parametric probability distribution function to describe the low flow sequences. There is also the problem of how to combine the individual errors for all the durations considered into a single measure for the assessment of Wright's model. However, using the storage-yield relationship is a direct approach, much simpler and does not involve all of the uncertainties associated with g90. For the work involved in the derivation of the storage-yieldreliability relationship, a computer programme has been developed using the Excel spreadsheet package (Nawaz, 1996).

Sensitivity analysis The storage-yield analysis performed previously assumed an initially full reservoir and also ignored direct evaporation fluxes from the reservoir surface. In this section, the impacts of these assumptions are briefly examined. If the impacts are negligible, then the use of the simple simulation approach to reservoir-yield-analysis for the purpose of model validation is justified, thus adding to its attractiveness. There is no doubt that considerations of the initial state and evaporation introduce additional complexities into the analysis and require additional information to implement, both of which may deter analysts from applying the test for evaluating the performance of data reconstruction models. For the sake of brevity, only the results based on three of the ten smaller catchments are reported: the larger Thames to Teddington catchment has been excluded, because its long record should ensure that the starting state of the reservoir has no significant effect on capacity estimates (McMahon & Mein, 1986). It is generally recognized that the starting state of a reservoir has an impact on the required storage capacity: the higher the deficit at the start of the simulation, the

128

higher will be the reservoir capacity required to meet the demand for a given reliability. This led McMahon & Mein (1986) to suggest the use of at least 100 years of data record for reservoir analysis in order to eliminate the impacts of the initial reservoir state. However, whether or not the initial reservoir state has an impact will depend on the "wetness" at the beginning of the inflow data record. If the beginning of the record is relatively wet, i.e. much higher than the mean, and precedes the critical period in the record, then the effect of any initial deficit on the subsequent reservoir states will be short lived; the resulting reservoir capacity will also be unaffected. If, on the other hand, the beginning of the record is dry, then this will probably lead to an extension of the critical period, causing the required storage capacity to increase, particularly at high yields. In general, the results of our simulations confirm such tendencies (Nawaz, 1996). As for the relative errors, Table 2 shows that there was little impact as a result of varying the initial reservoir state. This is probably due to the fact that whenever there is an impact, both the capacity estimates from the reconstructed and observed data records are affected; these effects then cancel out when the relative error is calculated. For evaporation, an average area-storage relationship based on data of surface area and the corresponding capacity for 21 reservoirs in England and Wales were used (Fig. 6). Although the scatter in Fig. 6 is indicative of a nonlinear relationship as expected, we have nonetheless approximated this by a linear relationship to enable its incorporation into the behaviour approach via equations (4)-(5) (see also Loucks

Table 2 Effect of the initial reservoir state on the relative error (%) of capacity for 98% time-based reliability (reservoir surface evaporation flux ignored). Catchment Yield = 30% Initially Initially full W full -26 -26 30 30 n/a n/a Yield = 60% Initially Initially full VA, full -38 -39 35 35 -40 -40 Yield = 90% Initially Initially full M full -24 -17 21 21 -22 -28

129

Wj (N

TJ-

os vd ^ o ri

rt

3

ON

wi rt rn <N

ES

^o <M - H

8 >< O

^ "* * " 1

oo ON iri c-i t ^ oo

l!

n * h

f at

(*

'g o

II

fui

"

> >

6? O VO II

2 >

a

tu

> O ^

ving aratio

en od M

,P c

<

Q/

3 ta

^ O

if

io *

in vd

w

fi

s H

3 S S O H ^

130

et al., 1981). Consequently, a linear regression equation was fitted to the scatter of points in Fig. 6, giving a particular form of equation (4a) as: A, = 0.0815, + 0.055 (R2 = 0.871)

2 6 3 2

(8)

where A, is the surface area (km ); S, is the storage (x 10 m ); and R is the coefficient of determination. Monthly open water surface evaporation data for the catchments were taken from Wright (1978). The effect of the inclusion of evaporation is to cause the storage capacity to increase for given yield and reliability (Table 3). This increase is large at high yields because it is here that the exposed surface area of the reservoir is also greatest. Assuming that the employed areastorage relationship is valid and assuming that the evaporation is not compensated for by any direct rainfall on the reservoir surface, which is likely to be the case during the critical period, this result further underlines the need to include explicitly surface evaporation fluxes in storage-yield analysis, even for reservoirs in temperate climates like England and Wales. However, because the inclusion of evaporation affects all the reservoir capacity estimates, whether based on the observed or reconstructed data records, there has been little effect on the relative error of storage (Table 4). The practical lesson here is that, while direct evaporation may be ignored when assessing the performance of a streamflow reconstruction model using storageyield analysis, the same may not be true where the objective of the storage-yield analysis is to plan a reservoir scheme or to investigate the influence of environmental factors such as climate change on a reservoir system. This is significant because most climate change impact studies appearing in the literature, while they have considered evaporation when perturbing inflow series, have often ignored reservoir surface evaporation fluxes, which this study implies could be significant.

Table 4 Effect of reservoir surface evaporation flux on the relative error (%) of capacity estimates for 98% time-based reliability (reservoir initially full). Catchment Ouse (Denver) Thames (Eynsham) Wensum (Costessey Mill) Yield = 30%

Ignoring evaporation Allowing for evaporation

Yield = 60%

Ignoring evaporation Allowing for evaporation

Yield = 90%

Ignoring evaporation Allowing for evaporation

-26 30 n/a

-29 28 n/a

-38 35 -40

-40 39 -41

-24 21 -22

-23 19 -19

DISCUSSION AND CONCLUSIONS The above results have reinforced the need to include criteria directly related to the intended use of reconstructed data for the verification of streamflow data reconstruction models. While the calibrated Wright's catchment models have performed reasonably well in preserving basic streamflow statistics such as the mean, standard deviation and skew of the data, they have not been as successful in reproducing the low flow sequences which are relevant for storage-yield analysis. This failure has

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment \ 31

translated into large errors (i.e. relative errors > 10%) in the derived storage-yieldreliability relationships and the g90 estimates. To illustrate the practical significance of the large relative errors associated with the storage-yield relationships, the results for the Thames to Teddington catchment are examined further. This is the largest of all the eleven catchments with an area of 9870 km2. It also has the longest historic data record, thus making it the ideal record to use for the purpose. Table 5 shows how reservoir system reliability is affected as a result of the errors. For example, consider a reservoir capacity obtained for the 98% time-based reliability using the observed data record. This reservoir will fail on average 12 months in a 50-year (600 months) period. However, if the same capacity had been estimated with the reconstructed data, this will only translate to a reliability of 94.75% for a yield of 0.5 of the mean flow, which is equivalent to additional 19 months of system failure. Similarly, the storage for 98% volumetric reliability based on the reconstructed data only managed a reliability of 96.34% with the historic record. Although this is not as low as the implied time-based reliability, it is nonetheless a significant increase in the total water shortage. The results for other yields and reliabilities presented in Table 5 show a similar trend. So what lessons could be learned from the foregoing? First, if the Environment Agency (EA) is going to commission further studies to calibrate Wright's model for more catchments in England and Wales, then it is important to include the storageyield relationships among the criteria for model verification. Secondly, although this is beyond the scope of the current study, it is also necessary to look at other streamflow extension techniques, to see if these offer better predictions of streamflow than Wright's model. One possibility is to use time-series analysis techniques (Salas, 1993); such methods not only extend the data, but they are also able to provide several realizations of flow data of any length which is important when studying the long-term performance of water resources systems (Vogel & Stedinger, 1988).

Table 5 The impact of errors in storage capacity on reliability. Volumetric reliability: required rel. = 9 8 % required rel. == 97% required rel. = 96% R* (%) R* (%) R* (%) 96.34 95.24 94.08 50 95.37 96.39 94.28 60 95.40 94.29 96.61 70 Time-based ;reliability (required = 98%) A/* R< AF* Yield AF* (% AMF) (months) (%) (months) (months) (%) (%) 19 94.75 19 93.75 93.12 17 50 95.20 17 94.20 17 93.48 15 60 95.56 15 94.84 13 93.30 16 70 * Actual volumetric reliability of the reservoir system as a result of under-design with reconstructed record. f Actual time-based reliability of the reservoir system as a result of under-design with reconstructed record. * Increase in number of failure months over 50-year design life as a result of under-design with reconstructed record. Yield (% AMF)

132

Hirsch (1982) (see also Salas, 1993) presented a number of regression-type approaches for extending streamflow at a short-duration site using streamflow data at an adjacent long-term station. In particular, he introduced the maintenance of variance extension, MOVE (types 1 and 2) techniques which ensure that streamflow variability, an important factor in reservoir analysis, is not reduced as would normally happen in classical regression analyses utilizing least squares parameter estimation (Matalas & Jacobs, 1964). As a consequence, the simulated flows are able to preserve well the extreme events in the historic record, particularly when the regression analysis uses the log transformation of flows at both the short and long-term sites (Hirsch, 1979). The approaches in Hirsch (1982) rely on the availability of a nearby long-term station whose flow characteristics such as distribution shape, serial correlation or seasonality are similar to those of the short-term station. However, it is also possible to substitute a long-term rainfall data if a long-term flow data record is unavailable (Raman et al., 1995). For the latter, Raman et al. (1995) demonstrated that performance was enhanced by twelve monthly regressions, rather than a single regression analysis which ignores the seasonality in both the flow and rainfall. Apart from using statistical methods, there are also several conceptual rainfall-runoff water balance models which have been successfully applied in simulating monthly flow records for reservoir analysis. A comprehensive review of these is available in Xu & Singh (1998). There is also the issue of ungauged sites: Wright's model cannot be used at ungauged sites because it requires at-site streamflow measurements for its calibration. Xu & Singh (1998) document studies in which conceptual water balance models have been used to simulate monthly flows at ungauged sites via a two-stage process: first calibrating the conceptual model at a number of gauged sites; and secondly relating the calibrated parameters to easily measurable catchment characteristics by regression analysis. The regression equations are then used to estimate the model parameters at ungauged sites for the model run. Hirsch (1979) also presented the simple area-ratio and regional statistics methods which could be applied to reconstruct flows at ungauged sites. In particular he found that, when applied to gauged sites, the regional statistics method performed as well as the log-log regression method in simulating low flows. This fact, coupled with the knowledge that the methodology could be used for both gauged and ungauged sites, makes the regional statistics method worthy of further investigation for the English catchments. The regional statistics scheme described in Hirsch (1979) is based on calibrating regression equations for the mean and standard deviation of flows in each month of the year, using catchment physical and climatological characteristics as independent variables, e.g.: V = a A1' S L'' S; E< F* (i>-20.0)* 7*4-2 T"' Sqn (9)

where a, b, c, d, e, f, g, h, k, m and q are regression coefficients; V is the dependent variable (i.e. monthly mean or standard deviation); A is catchment area; S is main channel slope; L is main channel length; S, is the percentage of area in lakes and ponds (plus 1%); E is basin elevation; F is percentage of basin forested (plus 1%); P is mean annual precipitation; 7242 is 2-year, 24-hour maximum rainfall intensity; r, is the minimum January temperature; and S is annual snowfall. Actual monthly

Monthly runoff estimated by a rainfall-runoff regression model for reservoir yield assessment

\ 33

flows at the individual sites can then be obtained using data from a long-term, benchmark station in the region: Ytt=Yim+Yis{-^) (10)

where Yim is the regional estimate of the mean flow at an ungauged site for month / (equation (9)); Yis is the corresponding standard deviation (equation (9)); M is the mean flow for month / at the base station in the region; 5 is the corresponding standard deviation; Xtj is the observed flow (month /, year j) at the base station; and Yjj is the reconstructed flow (month /', year j) at the ungauged site. The Thames to Teddington station, with its long data record, is a good example of a base station which could form the basis for using equation (10). The main conclusions of the current study are as follows: - A comparison of basic streamflow characteristics, e.g. the mean and standard deviation, is not sufficient when examining the adequacy of data reconstruction models; derived information which pertains to the intended use of the reconstructed data should also be examined as part of the calibration exercise. The simulation experiments have shown that the relative errors of the storageyield relationship derived with some of the Wright's model reconstructed data are higher than the relative errors in basic streamflow statistics. This is an indication that further refinement of the model is warranted to ensure that it simulates adequately the low flow regimes of river flow which determine storage-yield characteristics. Since most instances of data insufficiency relate to ungauged sites situation, for which Wright's model is not applicable, there is a need to examine ways of reconstructing time series data at ungauged sites. We have recommended regionalization as one possible method of accomplishing this. The limited sensitivity tests carried out have shown that the main conclusions relating to the relative error in storage estimates are unaffected by the starting state of the reservoir or the lack of consideration of direct evaporation fluxes from the reservoir surfaces. This is not true, however, with regard to the actual capacity estimates where both these factors have discernible impacts, particularly at yields of 70% of the mean flow and above. Acknowledgements The authors wish to thank Dr P. D. Jones of the Climatic Research Unit, University of East Anglia, UK, for supplying some of the data. Thanks are also extended to all the other organizations and individuals who took the time to reply to requests for information. We would also like to thank the two anonymous reviewers whose helpful and constructive comments aided us in improving upon the substance and clarity of the initial draft of the paper. REFERENCES

Adeloye, A. J. (1990) Streamflow data and surface water resource assessment: a quantitative demonstration of the need for adequate investment in data collection in developing countries. J. Wat. Supply Res. Tec. (Aqua) 39(4), 225-236.

134

Adeloye, A. J. (1996) An opportunity loss model for estimating the value of streamflow data for reservoir planning. Wat. Resour. Manage. 10(1), 45-79. Burges, S. I. & Linsley, R. K. (1971) Some factors influencing required reservoir storage. J. Hydraul. Div. ASCE 97(7), 977-991. Drayton, R. & Lambert, A. (1995) Surface water yield assessment. R&D Note 313, Environment Agency, London. Carty, J. G. & Cunnane, C. (1990) An evaluation of some methods of determining storage-yield relationships for impounding reservoirs. J. Instn Wat. Environ. Manage. 4(1), 35-43. Gan, K. C. & McMahon, T. A. (1991) Sensitivity of reservoir sizing to evaporation estimates. J. Irrig. Drain. Engng ASCE 117(3), 324-335. Hirsch, R. M. (1979) An evaluation of some record reconstruction techniques. Wat. Resour. Res. 15(6), 1781-1790. Hirsch, R. M. (1982) A comparison of four streamflow record extension techniques. Wat. Resour. Res. 18(4), 10811088. Jamieson, D. G. (1994) Eureka EU 487: a decision-support system for integrated river-basin planning. BHS Occasional Paper 4, British Hydrological Society, London. Jones, P. D. (1983) River flow reconstructions and their use. Part 1. Wat. Services 57, 528-551. Jones, P. D. (1984) River flow reconstruction from precipitation data. /. Climatol. 4, 171-186. Jones, P. D. & Lister, D. (1995) Extended flow records at key locations in England & Wales (Phase 1). Environment Agency, London. Klemes, V., Srikanthan, R. & McMahon, T. A. (1981) Long-memory flow models in reservoir analysis: What is their practical value? Wat. Resour. Res. 17(3), 737-751. Loucks, D. P., Stedinger, J. R. & Haith, D. A. (1981) Water Resource Systems Planning and Analysis. Prentice-Hall, Englewood Cliffs, New Jersey, USA. McMahon, T. A. & Mein, R. G. (1986) River and Reservoir Yield. Water Resources Pubis no. 138, Littleton, Colorado, USA. Matalas, N. C. & Jacobs, B. (1964) A correlation procedure for augmenting hydrologie data. US Geol. Survey Prof. Pap. 434-E, E1-E7. Nawaz, N. R. (1996) Assessing the adequacy of Wright's model reconstructed streamflow data for reservoir storageyield-reliability analysis. MSc Thesis, Heriot-Watt University, Edinburgh, UK. Pretto, P. B., Chiew, F. H. S., McMahon, T. A., Vogel, R. M. & Stedinger, J. R. (1997) The (mis)behaviour of behaviour analysis storage estimates. Wat. Resour. Res. 33(4), 703-709. Raman, H., Mohan, S. & Padalinathan, P. (1995) Models for extending streamflow data: A case study. Hydrol. Sci. J. 40(3), 381-393. Salas, J. D. (1993) Analysis and modelling of hydrologie time series. In: Handbook of Hydrology (ed. by D. R. Maidment), Chapter 19. McGraw-Hill, New York. Savic, D. A., Burn, D. H. & Zrinji, Z. (1989) A comparison of streamflow generation models for reservoir capacityyield analysis. Wat. Resour. Bull. 25(5), 977-983. Srikanthan, R. & McMahon, T. A. (1980) Stochastic time-series modelling of arid zone streamflows. Hydrol. Sci. Bull. 25(6), 423-434. Vogel, R. M. & Bolognese, R. A. (1995) Storage-reliability-resilience-yield relations for over-year water supply systems. Wat. Resour. Res. 31(3), 645-654. Vogel, R. M. & Stedinger, J. R. (1988) The value of stochastic streamflow models in overyear reservoir design applications. Wat. Resour. Res. 24(9), 1483-1490. Wright, C. E. (1978) Synthesis of river flows from weather data. Tech. Note no. 26, Central Water Planning Unit, London. Xu, C.-Y. & Singh, V. P. (1998) A review on monthly water balance models for water resources investigations. Wat. Resour. Manage. 12(1), 31-50. Received 1 April 1998; accepted 11 August 1998