Question 6 of 7: Frequency and Probability Analysis of Precipitation and Floods

Question 6: Frequency and Probability Analysis of Precipitation and Floods (20 marks)

Question text not reproduced: the examination questions are © Engineers and Geoscientists BC. Open the official past paper (linked at the top of this page) to read the question, then follow the worked solution below.

(i) Using IDF curves in design (6 marks)

An intensity–duration–frequency curve is the compressed statistical summary of a station’s short-duration rainfall record. It is built by extracting, for each year of record and for each of a set of standard durations (5, 10, 15, 30 minutes and 1, 2, 6, 12, 24 hours), the maximum average intensity observed over that duration; fitting an extreme-value distribution to each annual-maximum series; and plotting the fitted quantiles as intensity against duration, one curve per return period. In Canada these are published by Environment and Climate Change Canada for each principal climate station, and the IDF_CC tool provides climate-change-adjusted versions.

The curve is used in design in three related ways. First, and most directly, it supplies the design intensity for a peak-flow calculation. The designer computes the catchment’s time of concentration, selects the return period appropriate to the facility and its consequence of failure, enters the chart at a duration equal to the time of concentration, reads the intensity from the curve of that return period, and uses it in the rational method,

with i in mm/h, A in hectares and Q_p in m³/s. The reason the duration is set equal to the time of concentration is the core scaling argument: a storm shorter than that never has the whole catchment contributing at once, and a storm longer than that delivers a lower intensity, so the time of concentration is the duration that maximises the peak.

Second, the curve is fitted to an analytical form so it can be embedded in software. The usual expression is

where t_d is the duration and a, b, c are coefficients fitted separately for each return period at each station. This is what “scaling short-term precipitation data” means in practice: the relation lets an intensity be interpolated at any duration between the tabulated ones, and it is the form in which IDF data enter a stormwater model.

Third, the curve generates the design storm hyetograph needed when a full hydrograph, rather than a peak, is required — for detention-pond sizing, for example. The alternating-block method takes the depths at successive durations from the IDF curve for the chosen return period, differences them to obtain the incremental depth in each time block, and arranges those increments around a central peak to build a synthetic storm of the required total duration. The Chicago storm distribution does the same by an analytical rearrangement. In both cases the resulting hyetograph is, by construction, consistent with the IDF statistics at every duration.

Two cautions belong with any use of an IDF curve. Its extrapolation to return periods much longer than the record length is a statistical projection, not an observation — a 100-year quantile from a 30-year record carries wide confidence limits. And the assumption that the point-station intensity applies uniformly over the whole catchment is only reasonable for small areas; for catchments above roughly 25 km² an areal reduction factor must be applied, because the probability that the peak intensity occurs simultaneously everywhere falls as the area grows.

(ii) Why probability distributions are used, and two key hydrologic variables (6 marks)

Hydrologic processes are the outcome of atmospheric and land-surface dynamics that are, in practice, not predictable in detail more than a few days ahead and not repeatable at all over the decades that an engineered work must survive. The engineer nevertheless has to answer a deterministic question — how large must this culvert be — and the only honest route from an unpredictable process to a defensible dimension is through probability. Four reasons make the statistical treatment necessary rather than merely convenient.

The physical system is too complex and too poorly observed to be modelled deterministically over design lifetimes, so the record is treated as a sample drawn from an underlying population, and the distribution of that population is inferred. Design decisions are inherently risk decisions: a structure cannot be built to survive every conceivable event, so a level of exceedance probability must be selected, and only a fitted distribution can convert that policy choice into a number. The record is almost always shorter than the return period of interest, so the distribution is the extrapolation device that permits a 100-year estimate from a 40-year record — with quantified confidence limits attached. And a distribution provides a common, transferable language: the same fitted parameters support regional analyses, permit the pooling of short records from similar basins, and allow risk to be aggregated across a portfolio of structures.

The two most important hydrologic variables characterised in this way are the annual maximum instantaneous peak discharge (m³/s), which is the variable of flood-frequency analysis and the direct input to the design of bridges, culverts, spillways, dikes and flood-plain mapping; and the annual maximum precipitation depth or intensity for a specified duration (mm or mm/h), which is the variable underlying the IDF curves of part (i) and the design storms used for urban drainage and for probable-maximum-precipitation studies. Both are extreme-value variables, which is why the distributions used to describe them are drawn from extreme-value theory rather than from the normal family.

Two further variables complete the picture in practice and are worth naming: the annual minimum n-day low flow, which drives water-supply reliability, effluent-dilution permitting and instream-flow needs, and the annual runoff volume, which drives reservoir storage–yield analysis. Low flows are conventionally treated as extremes of the opposite tail and are fitted with a Weibull or log-Pearson distribution.

(iii) Extreme-value distributions and the prediction of future floods (8 marks)

An extreme-value distribution is the appropriate model for an annual maximum because of a theoretical result rather than a convention. If the annual maximum is the largest of a large number of independent events within the year, then as that number grows its distribution converges to one of only three limiting forms — Type I (Gumbel), Type II (Fréchet) and Type III (Weibull) — combined in the Generalised Extreme Value distribution. The tail of an annual-maximum series therefore has a known asymptotic shape, and fitting one of these forms is not curve-fitting for its own sake but the use of a limit theorem.

The Type I (Gumbel) distribution is the classic choice for flood peaks:

with location parameter u and scale parameter α. Inverting it and expressing the result in the frequency-factor form used in practice gives the design quantile directly from the sample statistics,

where x̄ and s are the mean and standard deviation of the annual maximum series and T is the return period in years. The procedure is therefore: assemble the annual maximum series from the gauged record; check it for independence, homogeneity and the absence of trend; plot the ranked data on extreme-value probability paper using a plotting position such as Weibull, P = m/(n+1), or Gringorten; fit the distribution by the method of moments, L-moments or maximum likelihood; test the fit; and read or compute the quantile at the required return period, quoting confidence limits. In North American practice the log-Pearson Type III distribution is frequently mandated instead of Gumbel, because its third parameter accommodates the skewness that real flood series exhibit; the fitting is then performed on the logarithms of the peaks with a regionally weighted skew coefficient.

The return period is defined as the reciprocal of the annual exceedance probability, T = 1/P(X ≥ x_T). It is essential to state what it does and does not mean: it is the average interval between exceedances over a very long period, not a schedule. Successive years are treated as independent Bernoulli trials, so the probability that the T-year event is exceeded at least once during an n-year design life — the hydrologic risk — is

which for the 100-year flood over a 50-year design life is 0.39. That number, not the label “100-year”, is what should inform the choice of design standard, and it is why critical infrastructure is designed to much longer return periods than the nominal design life would suggest.

Beyond the magnitude of the peak, hydrologic analysis is used to predict the further characteristics of future floods in several ways. The frequency curve is combined with a design hydrograph shape — from a unit hydrograph, or from a regional dimensionless hydrograph — to give the volume and duration associated with the peak, which is what detention and reservoir analyses require. Routing that hydrograph through the channel and flood plain, as in Question 4, produces the stage, the velocity field and the inundation extent, which are the inputs to flood-plain mapping and to the design of protection works. Regional frequency analysis pools records from hydrologically similar basins to produce a growth curve that can be applied at an ungauged site and that greatly reduces the uncertainty of long-return-period estimates from short records. And for the very largest events, where statistical extrapolation loses meaning, the deterministic Probable Maximum Precipitation and Probable Maximum Flood are used instead — these are the standard for high-consequence dam spillways under the Canadian Dam Association guidelines.

Finally, the stationarity assumption on which all of this rests — that the historical record is a sample from an unchanging population — is now explicitly questioned. Land-use change within the basin, regulation by upstream dams, and a changing climate all violate it. Current Canadian practice is to test the annual maximum series for trend and change points, to use only the homogeneous portion of the record or to adjust the earlier portion, and to carry a climate-adjusted scenario alongside the historical-frequency estimate for any long-lived structure.