You are on page 1of 6
TECHNICAL PAPER Journal of the South African Institution of Civil Engineering. 44(1) 2002. Pages 20-25, Paper 484A rm ALEXANDER professor of the Department of Givl and Biosystems Engineering of the University of Preteia, He has writen two books anda number of papers end 2000 he was a member of the Unites Nations Scientific and asters, and in 2001 he as a mem ber ofan international team of experts appomtod to advise the ‘government of Mazar om (he davelooment ofa national flood management poly Statistical analysis of extreme floods WJ R Alexander In this paper it is demonstrated that widespreaa, severe floods are caused by Infrequent, but not rare, meteorological phenom ena, including tropical cyclones and cut-off low pressure sys- tems. The magnitude of the severe floods relative to the series of annual maximum floods at any one site can be readily deter- mined, but all direct statistical analysis methods seriously under-estimate their frequency of occurrence. This is demon- strated by the statistical analyses of wide area, severe flood- producing rainfall. The statistical analyses confirm that the widely used log Pearson type 3 distribution using conventional moment estimators, remains the preferred method for the st tistical analysis of hydrological and meteorological data, but that no direct statistical analysis methods can be used with confidence for return periods exceeding 50 years. Procedures Tor overcoming these difficulties are presented in an accompa- nying paper on the standard design flood. INTRODUCTION wie urgent need for alternative design flood estimation procedures has arisen ‘ror the damage caused by the devastating floods that occurred over almost the whole of southern Atrca during the period December 1999 through to Merch 2000. “Thess were described a6 the most severe humanitarian disaster experienced on the subcontinent. At one stage more than a mil. ton people in the two northern provinces of South Alrica nad no access to potable water. Hundreds of lives were lost, end tens of Iousands of people in Mozambique and ‘hundreds in South Altica were transferred to refugee camps. More than 200 bridges were either destroyed or severely damaged and several thousand kilometres of black-top and sravel roads were damaged. The estimated repair cost to the overland communication infrastructure in South Aftia and Mozambique was estimated to be more than R000 million ~ costs that these countries can il afford. The neavy rainfall was described in Dyson and Ven Heerlen (2001), and other aspects wore described in a num- ber of written presentations atthe confer tence on the floods held atthe University of Pretoria in May 2000 (Alexander 2000, Van tsjon 2000 and Van Bladeren 2000). Neither the sevesity nor the widespreed nature of these loads wat unexpected. In 1990 and 1991 Alexander and Van Heerden carried out studies on widespread floods in South Africa. Findings were detalled in the presentation The destruction of bridges by floods during the past 120 years. What went ‘wrong? (Alexander Van Hocrden 19913) “This was followed by a substantial three ‘volume research seport, Determination ofthe risk of widespread interipton of communica- tons due to floods, commissioned by me Department of Transport (Alexander & Van Heerden 1991), The handbook Flood incroloy for Southern Africa (Alexander 1990) was revised and incorporated in @ 60-page handbook, Flood risk rection measures (Alexander 2001), but at that time the determination of the frequency of accuronce of severe, wide. spread floods remained elusive, Other than these publications and dam safety legisla- tlon, there are no South African national {guidelines for design flood estimation proce- ‘hates. The development of all design flood est mation procedures begins with the direct sta- tistical analysis of recorded data at gauged site, Te purpose of statistic analyses Is to determine the flood magnitude frequency (O21) lationship atthe site. The reliability of alternative statistical analysis methods is ‘addressed in this paper. There are several ‘methods available for transfering ths infor. ‘mavion co ungauged sites. The development ‘of new, robust method for this purpose is ‘esribed in the accompanying paper on the standatd design flood. RESEARCH AND PRACTICE At the outset, it has to he appreciated that all floods are the result of the complex interac ‘ion of hydrological and meteorological processes that take place on a wide range of lume an space scates. None of the processes hhave numerically quantifiable upper limits, and itis not possible to model them accu rately elther physically analytically or stats: tically. Consequently the results of all meth- ‘ods for determining the Q-T relationship have a wide and unquannidable band of “uncertainty about them. Thief particularly the case in South Africa with its wide range ‘of dlimatic conditions and exposure to severe, lood-producing meteorological ys- tems, lack of understanding of this situa tion Is the prime cause of the hydrological tunder-design of structures exposed to floods {in South Africa. and consequently the unac- ceptably high fallure rate Designers should also be aware of the ‘wide gap between research and practice, par- ticularly research based ont ubservations in milder climates. Examples ofthis conceen are expresced in Pilgrim 1986 (Australia), Bee tal 1993 (Canada) and Alexander 1994 (South Africa), Although this is not as critical, in South Africa because research inthe feld (f food nyarology fas beer strongly user oriented for many decades, it sil prevails, ‘even atthe level of international guidelines. Joernaal van die Suid-Afrikaanse Instituut van Siviele Ingenieurswese, 44(1) 2002 A number of statistical analysis methods have been proposed in the hydrological Iiterature over the years. The log Pearson Type 3 distbution using conventional moment estimators (LP3/MM) is obligato- 1y forall federal agencies in the USA (US Interagency Advisory Committee 1982) ‘but was subsequently criticised by Wallis and Wood (1985) and others. 11s Une rec ‘ommended method in Australia {Institution of Engineers 1998) where McMahon and Srikanthan (1981) found that the LP3/MM distebution was appro: priate for the estimates of extreme flood In the UK, the five-volume Flood et: ‘mation hanalook (PEH) was published after a major research effort (Institute of Hydrology 1999). The nev, theoretically based generalised logistic distribution using L-moment estimators (GL/LM) was recommended for application in the UK with its mild climate, but the authors warned that the applicability outside the Uk had not been evaluated. The LP3/MM Alistribution was not among the alterna- tive aistibuctons tested, despte ts yenet- al use internationally Te authors of the Canadian gulde- lines (National Research Council 1997) hada sceptical approach to claims of advances in direct statistical analyses, Tey maltained that where data series are sufficiently long, itis often possible to demonstrate substantial non-homo- _geneities in the form of discontinsiies, trends, or long period cycles. They expressed concern that at times it seemed at probability distsibutions and etl ds of fitting theoretical distributions to samples come and go with fashion. In most cases, the assumption that statistical theory applies is questionable, and a cer tain degree of scepticism about the results lf elaborate statistical procedures Is war ranted, Robée ot al (1003) suggosted that coordinated international study be under. taken, In South Africa, Alexander (1990) produced calculation procedures for & suite of seven statistical distributions for desktop computer applications. The phi Josophy was to allow users to select the ‘most appropriate method for the specifi problem being addressed. This is stil the philosophy in the revised handbook {Alexandes 2001) and the upgraded sulte ‘of computer programs (Van Dik & Alexander 2001). One major change is ‘the substitution of the new three parame: ter GL/LM distribution for the five parameter Wakeby distribution using probabilty weighted mouient esimators (WAK/PWM), developed by (Houghton 1978). which did not live up to its expec tations STATISTICAL ANALYSIS METHODS All dict statistical analysis methods are data fitting procedures. In most methods Table 1 Comparison of the calculated 50-year floods for selected sites and statistical distributions (m9/s) Tver Sie TSIM GEVIPWM Guim Blyde willemsoord 525 S21 313 Bree Taree Be i 3H Vaal Sanderton 1.666 1538 1596 Kat Fort Armstrong oo a9 392 Pienaars Rliparit 38 68 BS Tongolo Grootdraar BOSS, To Tear Mkomazi New weir TAT SIF CLT Table 2 Comnoarison of the QT relatonshio for the Pienaar River at Return period [LPS | LNIMM ] EVIIMM ] GEV/MIMT ] GEVIPWM | GL/LM (years) 2 50 0 9 7 so |v 10 200 200 381 aa 256 | ise 20 475 78 392 374 ‘400 | 288 50 828 837 536 674 as | 435 100 Ti9s [121s F) ao | rons 200 test [a ast) 1050 Tanz] ons ‘he aoumption Is made that the Gata ane identically distributed, which in turn assumes thatthe annual flood peak maxl- ma are te result of a single set of annul, flood-causative mechanisms that vary only in their magnitude This is not even approximately vali for many South ‘Atican catchments, SINGLE SITE ANALYSIS Annwal flood peak maxima at 152 repre: sentative sites totalling 6 728 years of records were provided by the Department fof Water Affairs and Forestry, Many of the data sets were upgraded in October 2000 by the department and include historical maxima. Table 1 shows the values ofthe 50> year design flac forthe three candidate listributions for direct statistical analyses in South Africa. They have been chosen. to demonstrate the similarities and differ fences of me results produced by the mEee methods. The tesults of the analyses In the Blydle, Bree and Vaal rivers are in close agreement, However, the results ofthe GL/LM distribution are appreciably lower ‘than those for the LP3/MM distribution in the Kat, Pienaars and Pongolo rivers, and much higher than those of the LP3/MM distribution in the Mkomazt River The reason for the erratic perform- ance of the GL/LM distribution 1s tat the developers of the method adopted a pol- «of deliberately developing 2 procedure that was robust against outliers, (FEM Vol 1, p 34). This was achieved by using lin- ear L-moment methods for distribution ting whien, 1 was claimed, resulted 1n less extravagant extrapolation of single site analyss. The high outers were assumed to have an undesirable influence fn the flood magnitude-frequency rela tionship, but this is not the situation in South Africa where high outliers ate cette Journat of the South African Institution of Civil Engineering, 44(1) 2002 ‘ally umportant, a8 these are the Doous that cause the damage The consequence is that in South, Africa the GL/LM distribution produces seriously anomalous results a sites where there ae outliers or other regularities in tu data, Tus alustratey by the Getaled “analysis ofthe data from the Dienaare River at Klipdof. This data set is a 94 year long record from a 1 028 km? catchment in the interior of South Africa that is not vulnerable to exceptionally severe floods. Table 2 shows te calculated Q-T relation: ships for six distributions, The three add tional distbutions ate the log normal (LNIMM), extreme value type 1 (EVI/MM) and generalised extreme value (GEV/MM) distributions all using conven tHonal moment estimators. The 30-year value for the GL/LM is approciably less than that for the other distributions and Js only half the value of the LP3/MM dis tribution, which isthe recommended dis tribution for South Atrican applications. Te 50-year values forte dhnee extreme value distributions are in close agreement with one another, but ess than the two log-transformed distributions which are mutually consistent (the skewness coe ‘lent of the logs i close to 2ez0 for this data se), This isa good example ofthe range ‘of results produced by the diferent stais- ‘ical distributions at many South African sites that are not affected by extreme Hoods. Another point to note is that by Ucfinition, if flood excceds the calcul ‘ed SO-yeat flood there i a SOK probabil. ty that it will also exceed the 100-year flood. In this example the 10-year flood assuming an LP3/MM distribution has a value 45% greater than the 80-year flood, This can be compared with other regions fof the wood listed in Farquharson etal (1987): France, Germany and the Netherlands (6%), USA and Canada (9%), ‘Australia (369), South Africa and Botswana (87%). There isa very close simiacity between Australian and south 2 He Figure 1 Combine distributions on inearextrame value distribution acales Figure 2 Combine distributions on logarithmic-normal distribution scales cern African ratios, which are vastly differ. successive segments ofthe ranked data ent from those in more humid regions. Where there are curvature changes with The dimensionless growth curve on lin. Increase in magnitude, a inthis case, the car-entreme value probability scales i Lemont estimacors result in QT ela ‘ure 1 illustrates the differences betwoen tionships that are appreciably fess than, the fitted curves forthe GL/LM distebu: those derived from the more robust con tion (identified) and three other distibu- ventional moment estimators, which are tions at this site. Tis is confirmed in the less sensitive to these anomalies, plot on a logarithmie-normal probability Figures 1 and 2 also llustrate the gap scale in figure 2. The reasons forthe poor between theory a praca. Figure 2 Is performance of the GL/LM distribution at based on the theoretical assumption that this site are twofold. The fist s the influ- the annual maxima are each the maxi ence ofthe low outliers below the thresh mum value of a number of occurrences old of 10 ms in figure 2. These can be luring the year, Figure 2 is based on the subjectively adjusted upwards to the cor- practical observation thatthe set of data responding values of the statitieally neu polnts ie approxumately on a straight line tual LN/MM Bt. This e preferable to con ‘when plotted on logertheienormal Soring these values and applying condition-_prohahilty scales. Departures from theory al probability and retrofitting algorithms. are not readily apparent In figure 1, but ‘The second reason is more difficult t© the consequences of inappropriate theo accommodate in statistical analyses, In retical assumptions when applied to fevence; the moment estimators are ‘South aican caca are ready apparent 1m 200.mm | N/A 94000 | 69.000 131.500 Ae receiving > $00 mm | None reso | 14 400 no cea recetving > 700 mmm | None 1730 1.600 None Return petiod (years) z z 3 3 Figure 4 Generalised logistic distibution applied to the 18 regional stations Table 4 Widespread rainfall event classification (percentages) Rainfall] 100 mm|]> 200 ma > 300-mm]> 400 mm]> S00 mm]> GOO mm [Number oF Gis [a 4125 [ctase 1 [16 1780) cane las 22 Class 3 12 4 109 Chast 20 8 A 2 Class 5 [60 32 16 12 a 4 Class 6 [80 50 0 20, 13 a 0 Table § Event classification of the four systems described in table 3 Date Location Weather stem [Region [Class Tanuary 1981 | Laingsbure- Cutoff low ¥ 4 Pott Elizabeth Tansary 1084 | fartern lowweld | Tropiesleyclone [7 4 Sepeembe 1987 | _KwazuneNaal | _cuvott iow 3 3 February 1988 | Bloemfontein. | Botswana low mn 3 Kumberly axis ‘tems include topical cyclones generated Jn the watm Indian Ocean in the north east and cut-off low pressure systems pro duced by westerly moving systems to the south of the continent, The Botswana lows are caused by Inland tropical tem= perate wave interaction, Note the broad m4 similaity ofthe rainfall produced by these diferent systems, These events were well documented 1m publications by Estle (1981), Kovacs (1982), Du Plessis (1984), Kovdes et al (1985), Tajaarc (1985), Van Bladeren et al (1987), Du Plesis etal (1989), Triegaardt etal (1991) and Tennant and Van Heerden (1993), DISTRICT RAINFALL The next step was to determine the fe quency with which these systems may ‘occur anywhere in South Africa. The South arrican Weather Service's monthly trict rainfall database (93 districts) was studied and the maximum monthly #39 fall over the whole of South Africa for teach of the 79 years of record was identl- fied. The data were ranked and subjected to statistical analysis, The return periods based on average rainfall over the whole ‘of South Africa for the months durlng which the four events in table 2 occurred ate shown on the bottom line ofthe table. Only the February L988 floods had ‘return period greater than 10 years on this basis For comparison, the rainfall associ cd with the February 2000 floods (Dyson. '& Van Heerden. 2001) hal a return period ‘of 17 years on this basis, Ths analysis shows that countrywide rainfall of ins magnitude is appreciably more frequent than ance in 50 yeas. SEVERE. WIDESPREAD RAINFALL CLASSIFICATION Tike next series of analyses was based on a ‘widespread rainfall algorithm developed by Van Heecden in Alexander and Yan Heesden (1991b). The regions are shown in figure 6 and the event clasifiation algorithm is shown in table 4 here the values are percentages of the open {accopted) stations during the month that recorded rainalls exceeding the specified rainfalls, Rainfall amounts used in the classification are the four-day totals. The properties ofthis algorithm are such that Uhey tend themselves to statistical analy The number of accurtences in South [Africa refers to the period from 1910 to 1989, Class 0 events may lead to signifi cant river flow but only localised flood: ing. Class 4 and 5 events result in disas- uous, widespread, severe flooding. Class 6 events have not been recorded in South Arca For comparison, the event classifica: tions for the four examples in table 3 are shown in table This analysis reinforces the conch son wached in the previous analyear that these widespread flood producing rainfall ‘events have return periods appreciably less than the SO-year return period tra tionally spectied in bridge desig, CLIMATE CHANGE The length of a number of South African data sets is now approaching 100 years ot gauged data, the Pienaars River data ‘escribed above being a good example. Tere ace also several sites where authen: ticated historic flood water levels have Joernaal van die Suid-Afrikaanse Instituut van Siviele Ingenieurswese, 44(1) 2002 ‘been recorded and the corresponding peaks calculated, Seventeen of the 906 values atthe 18 ites used in the regional attalyses above are histouical maxima, the caries being in 1847 in the upper Buffalo River near King William's town. ‘There are other sites where high flood levels have been recorded such as at Upington on the Orange River since 1874, and atthe Hankey mission om the Gamtoos River since 1867. The water lev els at Hankey ranked in onder of magni- tude occurred in 1867 Largest), 1932, 1971, 1847, 1916, 1905 and 1961. There {sno evidence in any of these records of a rogressive inerease or decrease in the magnitude of the Hood maxima. The damage ta ciil eng neering structures in recent yeas is due to the increasing numberof structures built Jn flood prone areas, and not an increase inthe frequency or Seventy of floods. ts ‘ko obvious that should there bean increase fn flood magnitudes of the order pactulat- ed by climatologists and others, the annu- al maxima would be lost in the Milky Way of values plotted in figure 4 above and will be statistically undetectable. ‘There mo jusuteauon for making any allowances for future climate change in design hard estimation procedures. CONCLUSIONS. International guidelines stress the need forthe application of sound engineering judgment in the determination and appli- cation ofthe design flood. These views have also been expressed by a number of experienced hydrologiss and practition- ‘ers i South Arca duing the past 30 ‘yeas. However, if hydrologists cannot ‘quantify their uncertainty. how can this uncertainty be accommodated in civil ‘engineering design? A solution is pro- posed inthe acompanyng paper on he standard des References Alexander WJ 81990. Fond tay fr sat fem ic, Sout Atean National Comittee ‘on tage Dams. Alexander, W JR 1994, Splway design Hoods — ‘esearch an prachice eg, nerationa Congyes on Large Dams, Durban. Imernational ‘Sammon on Large oan Alexander, W | R 2000, A season of ood. Cit Engen, Nay Alexander, W JR 2001, Hod is edton meus tres. Departmen or ew Engineer Unisray of Petri Alexandet, WR & Van Heotde,} 19914. The Gestion of bigs during the past 120 yeas. What went wrong? Proceedings, ATC Feseaich feu, vat ‘Alexander, WJ R& Van Heerden, J 19916. Decent peek of werent” aero a imma du los Department of anspor Research Report OAC 30/5, Bode, 8, Cava, G, Aska, Femi, ) Rasmissen, P2998- owas a systematic Sppreaeh to eamparing dstbutons used in ‘food frequency analysis ound of Hyde, 1H2s121-13e. sever Science Fublshes. Du Pies, DB 1988, Docanetaton of the ‘Marea 1951 fas te sax Cape Deparment of Water Afais Technical Report ean. 1 Pls, D8, Burger, CE. Dunsmore,$] & Randall, A 1989. Ducmeraton ofthe raat ods he Orage Re as Dera of Nate Als Teka evo Dyson 1.4 Van Heerden, J 2001. The heaey rain and Hoods over the datheasera te fro South afea during February 2000. Soa ‘Nea Journal of cece 1997 30-86, Farqunaron,r Ag Ofen, €3, nea, 1 Sune. V 1987. Campatson of flood fre {quency cures for many diferent regions ofthe ‘oe tn VP tng (ed), Reyna flo euency ‘hubs 225-286 Retde!Publshise ougnton, J 1979 urn ora paren: he ‘key station for modeling flood Hows. Water Resources Research, 1S IHOS-109. Ingtiunion of Faginees.Ausrala 1998. ‘soul anal andro ston of Engineers, asta Institute of Hydiology 1999, Fond timation hnubooktsttte of Hyerotogy, Ceowmast Gifford. United Kingdon, owas, £1982, Bocmemanon of te faraary TORT Ald the southern Cae. Department of Wate Aftais Techical Report ovkes 2 du Pes DB, Bracher PR, Dunn P ana Mary Gc, 1985, Document of ie ‘Journal of the South African Institution of Civil Engineering, 44(1) 2002 1984 Domoia fla, Department of Water Aas Retinal pore 123, Mevtahon, 1 A & Sokanthan, R 1981. Log Yearon Il dxtsbution =i i appa to ‘ood frequency analy of Ausallan streams? Journal of Hyd, 1981:139-147. ever Stent using Company, mse. ‘atlona Research Cour Canaéa 1997 yay of floes n Canada National Resear CCounet Camat Pugam, 0H 1986, Brdgng the gap Berween Na wnat a devon pears We Resear Resear, 2216591765, Tapa) 1985. Cuz lows in he sat ‘fron non. South Aca. Weather Bareay ‘rene Paper 14 “Tennant, W & Van Heerden, J 1993. The infu ‘ene of orographiy and loc sea surac teper- ‘ure anomalies in the development ofthe 1987 Natal floods. Sou Arica Joural of Slee ‘Tegandt, DO, Fetblanche DE, Van Heerden, J tang, MV 1888 The Natl load of Spun 1987 Weather Bure Tecnieat Paper 19 “iegaud, DO, Vn Heenden, J Steyn, PCL 1991. Anomalous pretation and ods daring eon 1988, Weather Bureau Technical Paper Us imeragency Advisory Committe on Water Data 1982 Bin (78 Gels fr der ‘ng lod low rune. US Department ofthe Inter, Geological Survey Van Bladven.D & Burger. CE 1989, ‘Documenta o he epee 198? Natal ows. Deparment of Water Aas Tectia por 139 \an ijn,» 20. ipod enaractenstcs at elected ioe and operation of reserve ring the February 2000 fod, Prac, {Conference on the Souter aracan Floods o Fabry 2000, Unies of Preto Van Bladeten, D & Van der Spey, D 2000. Te Fan 300 eas tn ing races, Contre oe Southern Aiea Hoods of ebrary 20 University of eto Van Dijk, M & Alexander, W J 2001 LUBTLODD fond nhs peas. Operators ‘anval Department of Cl Engineering, Uniersiy of Preto. Walls) 8& Wood, EF 1985, Relative accuracy fof the log Pearson Il grocedies Hyd En ASCE, Lug HOW 3066. 25

You might also like