Chapter 42 Frequency Analysis of Streamflow Data 424, General Frequency analysis of recorded stesmflow data is an ‘important flood-nnoff analysis tool. This chapter describes the role of frequency analysis and summarizes ‘the technicel procedures. EM 1110-2-1415 describes the procedures in groater detail a. Role of frequency analysis, (0) The trnditional solution to water resource plan- ning, designing, or operating problems is a deterministic Aelermicistic solution, a critical hyére- event is selected, This event is designated the desiga event. Plans, designs, or operating policies are selected to acconunadste that design event. Por example, the maximum discharge observed in the last 40 years may’ bbe designated the design event, A chermel modification may be designed to puss, without damage, this design event. If this design event is not exceeded in the next 1,000 years, tbe design may not be justified. On the other hand, if the dischorge exceeds the design event 20 times jn the next 30 years, the channel modification may be underdesigned, 2) A probabilistic solution employs principtes of Stites to quantity the isk that vasious bydrometearo- logical events will be exceeded. Risk is qurmied in terms of probability. The greater the risk, the greater the probability. Tan event is certain to occu, its probability 4s 1.00, If an event is impossible, ts probability ix 0.00, For food-runoff analyses, the probability of exeeedance ssually the primary interest. This is « measure of the risk that discharge will exceed a specified value, Decisions ae taken 0 that the risk of excecdance is acceplable For example, the channo] modification described. above sowld be designed for a discharge magninde with on annual exceedance probability of 0.01, Ia that case, the tisk is known and is acenonted for explicitly in the deci- sion making. 4, Definition of frequency analysis. (Q) The objective of streamflow frequency analysis is to infer the probability ef exceedance of all possible charge values {the parent population) from abserved dis- charge values (a sample of the parent population). This process Is sccomplished by selecting a statistical model that represents the relationship of discharge magnitude EM 1110-24497 34 Aug 94 and exceedance probability forthe parent popuation, ‘The ‘parameters of the models are estimated from the sample. ‘he calibrated model, the hydrologic engineer can predict the probability of exceedance for a specified mag- nitude or the mageitude with specified exceedance proba bility, This magnitude is refereed to as a quantile. @) For convenience, a statistical model may be displayed as a Gequency curve, Figure 12-1 fs an exam- ple of a frequency curve, The magnitude of the event is the ordinate, Probability of exceedance is the abscissa For hydrologic ongineering studies, the abscissa commonly shows “percent chance exceedance.” This is cexceedance probability multiplied by 100, (B) x some sense, skequency analysis is a model- ‘ting problem similar to the presipitation-runoll analysis problem described in Chapter 8 I bath cases, a model must be seleoted to describe the desired relationship, and the model must be calibrated with obeerved dat, €_ Summary of sireamfow Frequency analysis techn. ques. Techniques for selecting and calibrating steamtlow frequency models may be calegorized as graphical or sumerical With grophica! techniques, historical observe tions are plotied on spesiazed graph paper and the comves are fied by visual inspection, Nometical tot niques infer the characteristics ofthe model from satisties of Ge historical observations. ‘The proveduees or both ‘gapbical and numeriesl analysis ate presented in detail in EM [110-2-1415 and are summusized herein for ready reference. 12-2, Frequency Analysis Concepts 4 Data requirements, Statistical models of stceatn- flow fioquency are established by analyzing « sample of te variable of interest, For example, to establish a stats- tical model of annual peak discharge, the sample will be a series of annual peaks observed throvghout time, ‘The procedures of statistical anstysis require the following of ‘any time series used fa frequency analysis: 2) Dota must be homogeneous. That is, the data ‘must represent measurements of the same aspect of each event. For example, daily discharge observations should ‘not be combined with peak discharge observations, Furthermore, all sample points must be drawe from the same parent popuiation. For example, rain-flood data and snowmelt-flood data should not be combined if they esr be identified and analyzed separately. Likewise, dis: ‘charge data observed after development upstream should not be combined with predeveleprnent dat, 124EM 1910.24017 31 Aug 94 Magnitude 4.0 0.8 0.6 0.4 G2 0.0 Probabilty of Exceedance Figure 124, Frequency curve example @) Data snust be spatially consistent. AM data should bbe observed at the same location, Data observed at different locations: may be used 10 develop probability cstimates. However, these data must be adjusted t0 repre- sent conditions at « common location, (3) Time series must be continuous. Statistizal analy sis procedures require an urinferupied series. If obverva- font are missing, the misting valuct mast be estimated, for techniques for anslysis of broken records niust be used, 1b. Probability estimates from historicat data, (1) Streamflow protabitity is estimated from analysis ‘of past oscurence. The simplest model of the relstion- ship of streamflow mageitude ad probability is @ relative frequency miodel. ais made] estimates the probability of exceeding a specified magnitude ax the faction of time the magritnde was exceeded historically. For example, if tho mean daily discharge at given location excocds 122 80 ofs in 6,015 of 8,766 days, the reiative frequency is 9.68. The estimated probability of exceedance of 80 fs is 0.68, 2) Figwe 12-2 is a graphical representation of the relative frequency models of mean daily flow in Fishkill Creek at Beacon, NY. Such a plot is commonly referred to as a duration curve, ‘The abscissa of this plot shows “percent of time exceeded.” This equals relative frequeney rvulliptied by 100, 80 it is consistent with the term “‘parcent chinee exceedsnce.” G) The reliability of a relative frequency model impraves as the sample size increases: with an infinite semple size, telaive frequency exactly equals the proba- bility. Unfortunately, sample sizes available for streara- flow frequency analysis are small by scientific standards “Thus, relative lequency generally is uot a reliable estinaa- tor of probability for hydrologic engiveecing purposes.EM 4110-24417 31 Aug 94 SHRILL GREEK AT BEACON, NY T945THRS sot w E a | & z Sse 10h pa we Ey 90 2» 1 1 Or O08 PERCENT OFTHE EXCEEDED Figure 12-2. Graphics representation of relative fraquaney modal (4) The altergtive to the empirical relative frequency. model is a theoretical frequency model. With s theo- retical model the relationship of magnitude and. proba- bility for the parent population is hypothesized. ‘The rela- tionship is represented by = frequency distribution. A ‘cumulative frequecey disteibution is an equetion that defines probability of exceedance as a function of specie fied magnitude and one or more paxaueters, An inverse clistibution defies snagnitude as a function of specified probability and one or more parameters © Distribution selection and parameter estimation. Im cectain scientific applications, one distribution or another may be indicated by the pheromens of interest. This is not so in hydealogie engineering applications. Insteed, a frequency distribution is selected because it models well the data that are observed, The parameters for the model ate selected to optimize the fit. A graphical or numerical tcobnique can bo used to identify the appro- ‘plate distribution and lo estiguate the parsmeters, 42-3, Graphical Techniques. Some of the early and simplest mothods of frequency analysis were graphical techniques. ‘These techniques permit inference of the parent population charecteristi with a plot of observed magnitude versus estimsted ‘exccedance probability of that data, 1 4 best-tit ine is awn on the plot, the probability of exceeding various magnitudes can be estimated. Also, any desired quentiles ‘ean be estimated. Graphical representations also provide usefil check of the adequacy of a hypothesized diatibution, 2 Plotting position estiaates of probably (1) Graphicat techniques rely on plotting positions 10 estimate exceedance probability of observed events. The avedian platiing position estimates the exeeedance probs biligy as! (mw - 03) Wo oa - (21) where: = excendanes probability estimate for the mb largest event m= the order number of the event N= the qumber of events 42331 Aug 94 For exaimple, to estimate annual exceedance probability of anawal axiaimuor discharge, N= the nnmber of years of data, To express the results ag pereentchanes exceed ance, the results of Equation 12-T are multigliod by 100. @) Table 12-1 shows plotting positions for annual beak discharge on Fishkilt Creek, Column 4 of the table shows the discharge values in the sequence of oecurtence, Column 7 shows these same discharge values anranged in order of magnitude, Column 5 is the order number of cach event, Coluunn 8 shows the plotting position, ‘These plotting positions are values computed with Equation 12-1 tnd multiplied by 100. The values in columns 7 and 8 lus are an estimate of the peakcdischarge frequency isteibation, b, Display and use of estimated frequency curve. (8) The estimated frequency distribution is displayed fon a grid with the magnitude of the event as the ordinate aud probability of exceedance (or percent-chence exceed- ance) as the abscissa. ‘The plot thus provides a usefill ool fo: estimating quantiles or exceedanee probabilities. Specialized plotting grids are available for the display. ‘These grids aro consiructed with the abscissa soaled so a selected frequency distribution plots as a straight line. For cxample, a specialized grid was developed by Hazen for the commonly used normal frequeney distribution @) The specialized noumalprobability grid is a useful too] for judging the eppropriateness of the normal Table 424 ‘Annual Peaks, Sequential end Ordered with Plotting Positions (Fishkib Greek al Beacon, NY) Even Analyzed Ordered Everts Water Mosian Hoa Day Your how, chs Rank Year Flow, fe Plt Pos © ) eo o @ ® a @ 3 5 oy 2.200. + 1955 9.000. 287 2 2 1245, ‘470. 2 1956 8.280. 697 3 8 se 2.200. 3 1985 4.340. oF 3 8 sea 2570. 4 1968 3.0. 16.18 1 1 1849 3,020. 8 1953 3.20. 1928 3 ° 1950 420, 6 982 3.70. 2338 4 4 $056 2400, 1 192 3080. ons 3 2 1982 70. 8 sH49 3.020. ts 4 2% ses 3.220. 8 so4a 2970. ase6 ° ‘9 1956 4.700. 1958 2500. 9975 8 2» 4055 aco, " 1081 2490. 085 10 16 1955 8,280. 2 1945 2.280. ars 4 0 1957 t0, 9 17 2220. 5205 2 a 1997 2.500. “ 1350 2.0. ses 2 " 1983 +1960. 6 1059 11980 6025 4 ¢ ‘860 2.0 6 1980 780 baat 2 % 196 4340, v 1954 780. ea 3 8 ‘002 3,060. 16 1007 ‘680 284 a 2” $962 ‘720, © 146 70, race 1 zs 1904 1380, 2 1954 4380, pore 2 2 1965, ‘0, a 17 tao one 2 8 ‘1980 1,080. 2 1990 1210. 9 3 a 1867 5.580. 2 1986 1.040, ea 3 8 $968 3.630. 2 985. 580, ors 124listribation as & model of the parent population. If data drawn fiom « normally distributed patent population are assigned plotting positions using Equation 12-1 and are plotied on Hozen's grid, the points wil fall approximately fay a straight Hine, Ifthe points do not, then other the sample was drawn Srom a population with a diferent dis- ‘nibution or sampling variation yielded a nonrepresentative sample ) A specialized plotting grid has been developed also for nother commonly used fequency distribution, the log-normal distribution. Figure 12-2 is an example of such & grid. The values from cohmns 7 and 8 of ‘Table 12-1 are ploteed on this grid, ond a frequency curve is fitted. Ifthe data are truly daa from the distribution, (oF log-normal parent population, the points will fall ox EM 1110-24417 31 Aug 94 straight tine, The Fishkill Creek data, shown by Figure 12-3, do not fall on a straight line, 60 the assump- tion thal the pazent popuintion is a fog-normal distribution 5s suspect 12-4, Numerical Techniques Numerical techniques define the celationship between streamflow magnitude and probability with analytical tools, instead of the graphical tools, a. Steps of mumerical tecluigues, With aumerical techniques, the following general steps are wed to desive 2 frequency curve 10 represent the population (MeCuen ‘nd Snyder 1986): ‘ FISHKILL CREEK AT BEACON, NV 1945-1908 °° | ANNUAL PEAK FLOW, CF8 T 99 50988 0 A 0 + on oor PEROENT CHANCE EXCEEDANCE Figure 12-3. Log-normat probability grid 4125Enh 1110-2-1417 31 Aug 94. Q) Select a candidate frequency model of the parent population. Theee cistributions are commonly sed for fiequency analysis of hydrowetcorological data: the nor- ial distibuiion, the log-ronmat distibution, and the log Pearson type IT distribution, @) Obtain a sare, 3) Use the sample fo estimate the parameters of the model identified in step 1, (4) Use the modet and the parameters 10 estimate quautiles to construct the frequency curve that represents the parent population, 6 Numeriea! parameter estimation. (1) Parameters of a statistical model are commonly catimated from a sample with method-of anorsents estima: tors. The method-of-moments parameter estigytors are Aeveloped frome the following assumptions: (@) The sreanflow.probability selstionship of the prcent population ean be represented with a selected dis- inbution. The ssoments (derivatives) oF the distribution equation ean be determied with caleuis. One momeat is determined for each parameter of the distibution. The retuting expestions ace equations in terms of the param eters of the eistibution, (0) Moments of a sample of the parent population ean be computed munerically. The first moment is the meen of the sample; the second moment is the variance; the thied moment is the sample skew, Other moments ean be found if the distribution selected has more than three parameters, {©} The namerical moments ofthe sample are the best ‘estimates of the moments of the pareet population. This ssmumption pennits development of e set of simuancous equations, The distibuiion parameters are unknowe. in the equations, Solation yields estimates of the parameters (2) When the piramelers of the distribution sre esti- tasted, the inverse distribution defines the quantiles of the fiequency curve, Chow (1951) showed that with the ‘method-of- moments estimates, many inverse distributions ‘commonly used in bydiclogie engiveering could be wit- ‘en in the following general form: 1268 a G+Ks 29 where 1G, ~ the quantile with specified excardance probability p = the sample mean 'S = the sumple stondard deviation La frequency factor ‘The sample mean and standard deviation are computed with the following equations Ze, C8] woh where: 023) (2-4) Q, ~ observed event f N= number of events in sample (8) The frequency factor in Equation 12-2 depends ‘on the distribution selected. It is a function of the specie fied exceedance probability ond, in some cases, other population parameters, ‘The frequency factor fimetion can bbe tabulated or expressed in mathematical teams. For example, notmalstribution frequency feetors come- sponding to the excvedance pratabiliy p (0

0.5, (1 - p) is used in Equation 12-6, and the computed value of K, is uultipied by -1. (4) For the log:normal distribution, Equation 12-2 i writen as i, ~ Bes an tue Iogariti of Qy th dosined quantile X= mean of logarithms of sample $= standeed deviation of logasttms of somple K, = the frequency factor This frequency ‘ctor is the some as that used for the snormnal distribuiion. X¥ and S are computed with the fol owing equations: x Eee (29) ¥ 5 - [2o8.,- 2°] 4129) Ow here Q, = observed peak annual discharge im year # A’ = cumber of years ia sample Bor the annual peak discharge values shown in ‘Table 12-1, these values are as follows: X= 3.3684, and S= 02456, EM 1110-21417 31 Aug 94 & Recommended procedure for annual maxinun discharge, (@) The US. Water Resources Council (USWRC) (1967, 1976, 1977) recomnended the log Pearson type IIL distribution for onsial maximum steamflow frequency studies. This recommendation is followed by USACE. CCarrent gaidetines are peeseated in Bulletin [7B (USWRC 1980), @) The log Pearson type TI distribution models the frequency of logarithms of annual maximum discharge ‘Using Chow's (1951) format, dhe inverse log Pearson type I distribution is 210) 1X, © the logarithm of 0 he desied quantile x= ean of logarithies of sample standard deviation of logarithms of sample m " the Pearson frequency fketor X and $ are compoted with the Equations 12-8 and 12.9. ©) Fer this distribution, the frequency factor K is a function of the specified probability and of the skew of ‘the logarithms of the sample, The skew, G, is computed with che following equation: wee - a] cD yw -2 5) Ge For the values of Table 12-1, the skew computed with Equation 12-11 is 0.7300, @) The log Pearson type Tl Hequency factors for selected values of skew and exceedsnce probability are tabulated in Bullen 178 (USWRC 1981) and_ in EM 11102-1415. Alternavvely, an approximating 127eM 1910-24447 34 Aug 94 function can be used. IF the skew equals zero, the Pearson frequency factors equal tae normed distribution factors. Otherwise, the following approximation sug- gested by Kite (1977) can be used: Kom oleae og) - sat -(e whore k ~ G6. Analysis of special cases. (2), In hydroiogie engineering, applications, frequency ‘alysis of annual meximucn discharge is complicated by spacial cases. These include broken recores, incorplete records, zetoflow years, oullets, historical dota, and ‘small samples. Blletis 172 provides guidance for desl- ing with these eases. (2) AE 1 oF move years of data are missing fiom a time series of annual maximum discharge due to reasons rot related 10 Mood magnitude, the record is braken, For analysis, the record segments are combined, and the eom- bined record is analyzed as previously described. 3) AE data are nvssing because the events were too large (o record, too small to record, of the gnige was destroyed by a large event, the record ix incomplete. Any missing large events stould be estimated and the csti- ‘mates inckided in the time series. Misi are treated with the conditions! probabitity adjusorent seeommended for zere-flow years (The tog Peotson type HI ditibution f not sited (o sralysis of series which include oro low yeas, If ke saiple contains zero-flow years, the record is anlyzed Using the conditional probity procedure. With this procedure, the aubseries of nonzero peaks fs anlyzsd as described previously. The resulting frequency curve is a conditional fequeacy curve. ‘The exceednee fequencies Som this curve tre scaled by the relative fequeney of nop-zero flow year. The Tog. Pearson ‘ype TIT model parameters ore estimated for the upper portion of the ‘ouve, With these patimeters, a syetbetie frequeney curve 4s developed. Paragraph 3.6 of EM 1110-21415 describes the proces. 28 (5) Am outlier is an observation that departs signi cantly fom the tend of the remaining date. Procedures for treating oulliers require hydratogic and mathernatical judgment, Bulletin 17B describes one procedure for identifying high aed low outliers and for censosing the ata set. High outliers are treated a8 historical data if efficient information is available. Low ontliers are (reeted as zero-flow years, (6) Large floods ontside the systematically recorded tiene series may be used to extend that record, Thr proce: dure recommonded for analysis of these historiesl flows is 8 follows: (8) Assemble known historic peaks ord determine the historic record length. (b) Censor the systematic record by deleting all peaks tess than the miniwam historical peak. Estimate the model parameters for the remaining record. (©) Compute a weight withthe following equation: G-D (12-13) wD oe sexe: WF = the wie H = number of yeats in historic record 2 = umber of historie event LN = sumber of years in censored systematic tevord L = somber of zero low years, tow oat missing years exchuded from syst tic record (@) Adjust the model parameters with this woight Bauatons forthe adjusrents are prosented in Appendix 6 of Bulletin 17B (USWRC 1981), Compute te quantiles with these mocifid parameters ard Equation 12-10 (1) Small samples adversely effect the reliability of tstimates of the skew. This parameter is difficult ¢o est- ‘mate accurately from @ small sample. A more reliable estimate is obtained by considering skew characteristics of all available streamflow reconts in a large region. Atadopted skew is computed as & woighted sum of this regional skew and the skew computed with Equation 12-11. The weights chosen are a function af the sumple skew of the fogs, the saraple record length, the generalized skew, and the accuracy in developing the ageveralized values. The generalized skew can be deter- mined ftom @ unap included fa Bulletin 17B, or it can be determined fom detailed analyses if additional dats are availabte, (8) The impact of uncertainty due to stall sample size canbe quantified farther with the expected probability adjusiment. This adjuarment is based an the argument that the x petcent-chunes discharge estimate ‘made with a given sample is approsiaalely the median of alt estimates that wonld be made with successive samples of the sume size. However, the probability distibuson of the estunate is skewed, 50 the average of the samples exceeds the median. The consequence of this is hat if very luge nurrber of estimates of flood magnitude are made over a region, more x percentchance floods will oseur than expected on the average (Chow, Maidnent, sand Mays 1988). For exanple, more “100-year floods” ‘will ooetr in the United States annually than expected. Peregraph 3-4 of EM 1110-2-1415 describes how cither the probability associated with a specified magnitude or the magnitude fora specified probability can be adjusted to obtain a frequency curve withthe expected number of exceedances. Verification of froguency estimates. (2) The seliabilixy of frequency estimates depends on how well the prozosed model represenis the patent popa- lation, ‘The fit can be tested indireetly with a simple Braphical comparison of the filed model and the sample oF with a more rigorous statistical (est, The reliability can also be ilustrated with confidence limits. @) A grephical test provides a quick method for ‘verifying frequency estimates devived with rumerical procetiutes. The test is performed by plotting observed augaitude versus plolting-position estimates of exceedance probability. The postelated fieauency curve ‘With best-estimate parameters is plotted on the same grid. Goodness-of-fit is judged by inspection, as. desorbed previously. @) Because of the complexity of the log Peatson type HI distribution, no single specialized plotting grid is practical for this graphical test. Instead, the log-nocmal ‘tid is used to display data thought to be drawn from a fog Pearson type IIT distribution. ‘The iit is judged by EM 1110-24417 31 Aug 94 inspection. Figure 12-4 ifustates this, The observed peaks and plotting positions from columns 7 and 8 of Table 12-1 are ploited ete. Quantiles computed with Equation 12-10 se plotted on the seme grid, The estimated values of the terms of Equation 125 are X= 3.3684; 5 = 0.2456, and G = 0.700, The skew was ‘adjusted here with a regional skew. Tie computed fre- queney curve fits well the plotted observations. (A) Rigorous statistical tests permit quantitative judgement oF goodness of fit, These tests compare che theoretical disuibution with sample values of the celative frequency or cunmfative frequency fonction. For exam ple, the Kolmogoroy-Srnimov test provides bounds within ‘which every observation should Tie ifthe sample actually is drawn from the assumed disttibation, The test is sou- ducted a follows (Haan 1977}: (@) For each observation in the somple, determine the relative exceedance frequency, This is given by m/v, where m = the number of observations in the sample grester than o¢ egual to the observed magnitude, and A= the number of observations (2) For cack magnitude in the sample, determine the theoretical exceedance frequency using the hypothesized ‘model and the best estimates of the parameters, (©) For each observation, compute the difference in the relative exceedance frequency ané the theoretical exeeedanee frequency, Determine the maximum differ- nee for the sample, (@) Select an acceptable significance level. This is = measure of the probability that the sample is not drawe from the candidate distribution, Values af 0.05 end 0.01 tte common. Determine the corresponding Kolmogorov- Smimov (est statistic. ‘This statistic is 2 fonction of the sample size and the significance level. Test statistics ace tabilated or can be computed with the following equation (Loucks, Stedinger, and Harth 1981); c 244 2+ where: C= 1.358 for significance level 0.05, © = 1.628 for significance level 0.01 12.9EM 1110-24417 31 Aug 94 oe FISHKILL CREEK AT BEACON, NY 1245-1968 A - z 201 E ‘OBSERVED ANAUAL Pens coupes raravency conve wl mm 8S ee % 1a 0s PPEROENT CHANCE EXCESDANCE Figure 12-4. Plot for verification (©) Compare the maximum difference determined in sep © with the test statistic found in step a. If the value in sep o exceeds the test statistic, the hypothesized disc ‘ution eannot be accepted with the specified significance level (5) The cticbitty ofa computed frequeney curve can be illustrated conveniently by confidence limits plotied on the fiequeney gtid, Conthleace limits are established considering the uncertainty in etieating population mcan and standard deviation from 4 soll sample, For 2on- venieace, Appendix 9 of Bulletin 178 (USWRC_ 1981) includes table of equency factors that pemsit definition of L pereent (0 99 percent confidence limits. ‘These fe> quency ficiors ates function of spocifed excocdence ‘probability and sample size. As the sample size inveases, the limits nartow, inizatng incceased tlibiity. (© Figuee 12-5 shows the Sand 95 percent confi- ence limits for the Fishkill Creek frequency curve. The 1240 probability is 0.05 that the true quantile for a selected cxevedanee probability will exeeed the value shown on the S percent curve. ‘The probability is 0.95 that the tue quantile will exeeed the 95 percent-curve value and only (0.05 that it willbe Jess than the 95 perceat carve, 412.5. Spoctal Considerations @ Mixed paputations. In ceria. cases, observed stecamfow is thought to be the resull of two oF more independent hydrometeorologicat conditions. Tie sample is refered to as a mixed-population sampte. For example, the spring streancflow in the Sacrarieata River, CA, isthe resuft of both rains and shownell, For these cases, the data sxe segregated by cause price to analysis, i possible. Each set can be analyzed separately to determine the approprists ditibution and parameters. The resuling fequeney curves ate then combined using the following equation to determine probability of union:EM 4110-24417 ot uy 94 FISHKILL CREEK AT BEACON, NY 1945-1968 [ * || 107 PP OBSERVED ANNUAL PEAKS: i Secret ve > Sea menue cue See eee we | ! | Lt oo * a Figure 12:8, Frequency curve with cantidence limits Pa +P BP, 412-5) wheres P. = annual exeeedance probability of combined populations for a selected quantile , = annual exceedance probabifity of same magaitude for sample 1 Py = annual exceedaace probability of same ‘magnitude for sample 2 ‘This assumes Cat the series are independent, Otherwise, ‘coincident frequency analysis must be used, 2 Coinckdeat frequency analysts. in some planning, designing, or operating problems, the hyérometcorlogical event of interest is a function of two or more random Ihydrometeoralogicsl events (1) For example, discharge at the confluence of teo streains is a function of the coincident discharge in the ‘wibutary streams, The adjective of coincident frequency’ analysis is to estimate the Gequency distribution of the result if the frequency di known. The specific tec ‘mathematical form of the function relating the variables, ‘Benjamin and Cornell (1970) describe « variety of solu- tions, including analytical closed-form solutions and ‘Monte Carlo simulation, @) Ie hydrologic engineering, the variable of interest offen is the sum of components. fn that ease, the fie: quency distribution of tie sum can be found through conditional probability concepts. For itfusteation, consider the total discharge downstream of a confluence, Q>, is computed as the stm of tributary discharge Q, wred tributary discharge Q,. The trequency of Q, and Q, are established using procedures described previously. Roughly spenking, the probability thai Q; equals some neatEM 1110-21417 31 Aug 94 specified value, gp is proportional to tle probability tit 2, cquls a specified value, gy, times a factor proportional to the probability that Q, equals gp - 4). This prod ie summed over all possible values of Q,. To dovelop a frequency curve for the sum, the process is repeated for all possible values of Q,, Chapter 13 of BM 1110-2-14i5 prescate 1 deailed example of coincident frequency aualy- sis, € Regional frequency anclysis, (1) Methods of frequency analysis described _pre- viously ia this chepter eppiy (0 data coliected at a single site. IF large sample ig available at that site, the result ing frequency snalysis may be sufficiently reliable for planning, designing, or operating civil-works projects. However, samples commonly are small. In fact, i is nol ‘arusual that eisk information is required st sites for which no data are available, Regional Sequesey analysis tech- ‘niques may tse used to develop this information @) Regional frequency enelysis procedures. relate parameters of a streamflow-frequency model to catchment characteristics. Briefly, the following general steps are followed fo derive such 2 relationship: (©) Select long-record sites within the region, and collect stroumflaw data Zor those sites, (B) Select an approprintc distribution for the data, and estimate the parameters using the procedures ceseribed herein, (© Sclect catchment characteristics that should cor- relate with the parameters, Measute or observe these characteristics for the longetecord sites. Typical charac- teristics for streamflow frequency model parameters 1242 include the following: contributing drainage a4ea, stream Tength, slope of eatchment or main channel, sueFice stor age, mean sural rainfall, number of rainy day's snoually, infiluetion eharacteristes, and impervious area, (@) Pertoom a regression aualysis to establish predic- tive equetions. The dependent voriables in the equations fare the frequency model parameters. The independent variables are the calchment characteristies, (3) EM 11102-1415 provides sd establishing regiousl equations. nal guidance in 4 Frequency of other hycronsteoroiogcat ghee nomena, The procedures described for dischorge- frequency analysis apply to malysis of ther Iydrometcorlogical pheaoutena, The same gererl steps resented im parngreph 124 ae followed, For case the variable of interest is sesmflow volvine, eather than discharge, the time series will ben sequence aF volnmes foe a specified dination. “The procedures fer selecting calibeating, end verifying fequeney model aro the same ss preslounly described Volume-fequeney mel preiptation-depth-dra- fionseauency anaiyses, "These analyses present some uniqoe probiens.Beeeuse of the soil samples fiom which parameters must Be estimated, the sot of frequency curves. for various durations may ‘be inconsistent. For example, the Iiry vole should not exceed the 3-day volume forall probabilities, Yet, for a small samp, the conputed curves may not follow this ras, To overcome thisy the compuied curves may bo “voted,” ested ty inspection of plots. Altemavely, the statsea! model patametsis can be adjusted to mainiin cosisteey.Pae- friph 3c of EM. 11102-1415 desis a. typical mooedure.Page intensionatiy left blank