Question 6 of 7: Frequency and probability analysis of precipitation and floods

Problem 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.

Part (i) — What flood-frequency analysis predicts, and what it does not (7 marks)

Flood-frequency analysis is the statistical procedure that converts a record of observed annual maximum discharges into an estimate of the discharge associated with any specified probability of exceedance. Its use in prediction is best explained by being precise about the two halves of the question, magnitude and timing, because the method treats them very differently.

Predicting magnitude. The procedure is the one set out in Problem 4(iii): extract the annual maximum instantaneous peak for each year of record, check the series for independence, homogeneity and trend, fit an extreme-value distribution by the method of moments or L-moments, and evaluate the quantile for the return period required. Chow’s frequency equation $Q_T = \bar Q + K_T s$ delivers the magnitude, and the standard error of the fitted quantile delivers its confidence band. The magnitude estimate is genuine prediction in the useful sense: it says that a structure designed for $Q_{100}$ has a one percent chance of being exceeded in any year, and the design can be tested against that. Its reliability depends on record length, on the appropriateness of the distribution in the tail, and above all on the stationarity of the record; a basin that has urbanised, been logged, been regulated by a dam, or is experiencing a changing snowmelt regime under a warming climate violates the identical-distribution assumption, and the analysis must then be performed on a record adjusted to a common land-use condition or replaced by continuous simulation.

Predicting timing. Here the method must be described honestly, because a frequency analysis cannot forecast the date of the next large flood. What it does provide is timing information of three other kinds. First, it gives the probability of occurrence in a stated period: the risk that the $T$-year event is equalled or exceeded at least once in $L$ years is $R = 1-(1-1/T)^{L}$, so a 100-year flood has a 22 percent chance of occurring during a 25-year project life and a 39 percent chance during a 50-year one, and the expected waiting time between exceedances is $T$ years. Second, it gives the seasonality of flood generation when the analysis is stratified by cause — in most Canadian basins the annual maximum arises either from spring snowmelt, often augmented by rain on snow, or from a summer convective or autumn frontal storm, and separating the record into these populations (a mixed-population analysis) both improves the fit and tells the operator which months carry the risk. Third, in combination with a rainfall-runoff and routing model it gives the within-event timing: the design quantile is used to scale a design storm, and the model then predicts the time to peak and the duration above any threshold, which is what reservoir operations and evacuation planning actually need.

The practical uses follow directly: sizing spillways, culverts and bridge openings to a regulatory return period; defining the regulatory flood plain, which in most Canadian provinces is the 200-year flood; setting reservoir flood-control allocations; establishing flood-insurance and land-use limits; and providing the design inflow for the dam-safety classification of the Canadian Dam Association guidelines.

Part (ii) — Why distributions are used, with one frequency and one distribution named (7 marks)

Hydrologic variables — rainfall depth, peak discharge, snow-water equivalent, low flow — are the outcome of atmospheric and catchment processes so numerous, so non-linearly coupled and so sensitive to initial conditions that they cannot be predicted deterministically beyond a forecast horizon of days. They are, however, remarkably stable in their statistical behaviour: the mean, variance and shape of the distribution of annual maxima at a station change little over decades, even though the individual values are unpredictable. Treating a hydrologic variable as a random variable and describing it by a probability distribution is therefore not an admission of defeat but the only formulation in which the question an engineer must answer can even be posed.

Four specific reasons justify the practice. First, design requires extrapolation beyond the record. A culvert must pass the 100-year flow, but the record is 40 years long; only a fitted distribution can supply the 100-year quantile, because the data themselves contain no event of that rarity. Second, risk must be quantified, not merely bounded. Engineering decisions trade the cost of a larger structure against the expected cost of failure, and that trade-off requires a probability, not a worst case. The economically optimal design is the one minimising the sum of capital cost and the expected annual damage $\sum p_i D_i$, and the $p_i$ come from the distribution. Third, distributions compress a record into a few estimable parameters, which allows information to be transferred: regional frequency analysis pools records from hydrologically similar gauged basins to estimate the parameters for an ungauged one far more reliably than a short local record could. Fourth, they make uncertainty explicit — the standard error of a fitted quantile, and hence the confidence interval on the design flood, is computable, which is what allows a professional engineer to state the reliability of a design rather than assert it.

One frequency commonly used: the annual exceedance frequency, expressed as the return period. The return period $T$ is the reciprocal of the annual exceedance probability, $T = 1/P$, so a one percent annual exceedance probability is the 100-year return period. It is the standard currency of hydrologic design in Canada — municipal storm sewers to the 5- or 10-year event, major drainage systems to the 100-year, regulatory flood plains to the 200-year, and dam inflow design floods up to the probable maximum flood. Empirically, frequencies are assigned to ranked observations with the Weibull plotting position $P = m/(n+1)$.

One distribution commonly used: the Gumbel, or Extreme Value Type I, distribution. Its cumulative distribution function is

with location parameter $u$ and scale parameter $\alpha$, estimated by the method of moments from $\alpha = \sqrt{6}\,s/\pi$ and $u = \bar x - 0.5772\alpha$. Its theoretical justification is strong: extreme-value theory shows that the maximum of a large number of independent, identically distributed variables with an exponential-type tail converges to this distribution, and an annual maximum is precisely such a maximum taken over the storms of a year. Inverting it gives the frequency factor quoted in Problem 4(iii). The other distribution in universal engineering use, and the one prescribed by much North American practice for flood flows, is the log-Pearson Type III, in which the logarithms of the annual maxima are fitted by a three-parameter gamma distribution; its third parameter, the skew, gives it the flexibility to match records that the two-parameter Gumbel cannot, at the cost of requiring a regional skew estimate when the record is short. For rainfall depths the Gumbel and the generalised extreme value distributions dominate; for low flows the Weibull and log-normal are used.

Part (iii) — Deriving an IDF curve, and two assumptions for urban use (6 marks)

An intensity-duration-frequency curve family, of which the printed figure is an example, summarises at one location the relation between rainfall intensity, the duration over which that intensity is averaged, and the return period. Its derivation is a frequency analysis performed separately for each duration.

1. Assemble a long record of short-interval rainfall at one station. A tipping-bucket or weighing gauge recording at 5- or 15-minute intervals is required, with at least 20 to 30 years of record; in Canada these are the Environment and Climate Change Canada short-duration rainfall stations, and the derived curves are published as the ECCC Engineering Climate Datasets.
2. For each standard duration, extract the annual maximum average intensity. The standard durations are 5, 10, 15, 30 minutes and 1, 2, 6, 12 and 24 hours. A moving window of each duration is passed over every year of record and the largest average intensity found in that year is recorded, giving one annual maximum series per duration. The maxima for different durations need not come from the same storm, and usually do not.
3. Fit an extreme-value distribution to each duration’s series. The Gumbel distribution is the standard choice in Canadian practice. Its quantiles are evaluated for the required return periods using $x_T = \bar x + K_T s$ with the Gumbel frequency factor, giving the intensity for each duration-return period pair.
4. Plot and fit a smooth analytical form. Plotting intensity against duration on log-log axes for each return period yields the descending family of the printed figure, near-straight because the underlying relation is close to a power law. A smooth equation is then fitted for use in design software, most commonly $$i \;=\; \frac{a}{(t+b)^{\,c}}$$ where $i$ is intensity in mm/h, $t$ is duration, and $a$, $b$ and $c$ are constants with $a$ increasing with return period. The two-parameter form $i = a\,t^{-c}$ and the Canadian convention of tabulating $a$, $b$ and $c$ per return period are both in use.
5. Evaluate the fitted curve for a design case. With assumed coefficients $a = 1000$, $b = 5$ and $c = 0.75$ for a 25-year return period, $t$ in minutes, the intensity for a 30-minute storm is $$i = \frac{1000}{(30+5)^{0.75}} = \frac{1000}{14.39} = \boxed{69.5\ \text{mm/h}}$$ and, applied through the rational method to a 12 ha urban catchment with a runoff coefficient of 0.65, $$Q = \frac{C\,i\,A}{360} = \frac{0.65 \times 69.5 \times 12}{360} = 1.51\ \text{m}^3/\text{s}$$ which is the peak flow the storm sewer must carry.

For ungauged locations the curves are regionalised, either by transferring a nearby station’s curve scaled by the ratio of mean annual maximum rainfalls, or from published regional maps.

Two underlying assumptions for urban applications.

Assumption 1 — stationarity of the rainfall record. The frequency analysis behind every point on the curve assumes that the annual maximum intensities are drawn from one unchanging population, so that the past record is a valid sample of the future. In an urban context this is now the most consequential assumption, because short-duration convective extremes are the rainfall statistic most sensitive to a warming atmosphere: the Clausius–Clapeyron relation implies roughly seven percent more atmospheric moisture per degree of warming, and observed sub-hourly extremes in several Canadian cities are increasing faster than the annual totals. A curve derived from a 1960–2000 record may therefore understate the intensity a sewer will actually face over its 75-year life. Canadian practice increasingly addresses this by applying a climate-change adjustment, for example with the ECCC and Western University IDF_CC tool, or by adding a stated allowance to the design intensity as several municipalities now require.

Assumption 2 — the design storm is uniform over the catchment, and the storm duration equals the time of concentration. An IDF curve is a point statistic derived from one gauge, but it is applied to an area. Using it directly assumes the intensity is uniform across the catchment, which is acceptable for the small urban catchments of a few tens of hectares to which the rational method is applied, but requires an areal reduction factor for catchments beyond roughly 25 km², since a convective cell rarely covers a large area at its peak intensity. The companion assumption, embedded in the rational method, is that the critical storm is the one whose duration equals the catchment’s time of concentration, so that the whole catchment contributes simultaneously; this is why $t$ is set equal to $t_c$ when reading the curve. A third assumption implied by the same reading — that the return period of the resulting flow equals the return period of the rainfall — holds only when antecedent conditions and losses are near-average, and fails for the imperviousness-dominated response of a fully urbanised catchment in dry conditions.