Question 4 of 7: Frequency and Probability Analysis of Precipitation and Floods

Problem 4: 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) — Four differences between the IDF method and the Rational Method (8 marks)

The premise of the question needs one clarification before the differences can be listed: the two are not competing predictors of peak flow. An IDF curve is a rainfall product; the Rational Method is a runoff model that consumes it. With that established, the four substantive differences are:

1. What each one outputs. The IDF relation returns a rainfall intensity in mm/h for a chosen duration and return period — a meteorological statistic derived from the gauge record, with no reference to any catchment. The Rational Method returns a discharge in m³/s for a specific catchment. The IDF curve therefore has no runoff coefficient, no area, and cannot by itself predict a flow.

2. How the frequency of the answer is established. The IDF curve’s return period is an observed statistic: annual maximum intensities for each duration are extracted from long records and fitted to an extreme-value distribution, so “25-year intensity” has a defensible empirical meaning. The Rational Method simply assumes that a rainfall of return period T produces a flood of return period T. That assumption is convenient and demonstrably imperfect, because antecedent soil moisture, storm movement and areal patterns all break the correspondence.

3. Treatment of time and duration. The IDF relation is defined across the whole range of durations, and the inverse relationship is its whole point: short storms are intense, long storms are not. The Rational Method collapses this to a single duration — the time of concentration — on the argument that the peak occurs when the entire catchment contributes simultaneously. It therefore produces a peak flow and no hydrograph at all, whereas the IDF curve can drive a full design storm and hence a complete hydrograph.

4. Range of valid application. IDF curves are valid wherever the gauge record supports them, for any catchment size, subject to an areal reduction factor for large areas. The Rational Method is restricted by its assumptions — uniform rainfall in space and time, a constant runoff coefficient, and negligible storage — to small, largely impervious catchments, conventionally under about 80 ha (0.8 km²). Applying it to a 100 km² rural basin is a misuse; the SCS or a unit-hydrograph method is required there.

Given. The IDF relation printed on the paper, $I_T = 170\,T_r^{0.24}/T_d^{0.69}$ with I in mm/h, T_r in years and T_d in minutes; a catchment of A = 50 ha with runoff coefficient C = 0.60 and time of concentration t_c = 20 min; design return period 10 years.

Find. The design intensity and the corresponding Rational-Method peak flow, showing how the two methods chain together.

1. Read the design intensity off the IDF relation. The Rational Method requires the duration to equal the time of concentration, so $T_d=t_c=20$ min: $$I=\frac{170\,(10)^{0.24}}{(20)^{0.69}}=\frac{170\times 1.738}{7.902}=37.4\ \text{mm/h}$$
2. Apply the Rational formula in metric form. With $A$ in hectares and $I$ in mm/h, the constant 360 converts to m³/s: $$Q=\frac{C\,I\,A}{360}=\frac{0.60\times 37.4\times 50}{360}$$ $$\boxed{Q = 3.12\ \text{m}^{3}\text{/s}}$$ This is the chain in full: the IDF curve supplies $I$, and the Rational Method converts it to $Q$. Neither step can be omitted.
3. Note how the answer scales with return period. Because $T_r$ enters only through $T_r^{0.24}$, moving from the 10-year to the 100-year event multiplies the intensity by $10^{0.24}=1.74$, giving 65.0 mm/h at the same 20-minute duration — independent of duration. A tenfold increase in rarity buys only a 74 % increase in intensity, which is why upgrading a minor storm sewer from a 10-year to a 100-year standard is far less costly than the return periods suggest.

Part (ii) — Determining the recurrence interval of a flood event (4 marks)

Given. An annual maximum series of instantaneous peak discharges for the river, n = 30 years, drawn from the Water Survey of Canada HYDAT archive; sample mean $\bar{x}$ = 450 m³/s and sample standard deviation s = 120 m³/s.

Find. The recurrence interval assigned to a ranked observed flood, and the magnitude of the 100-year flood.

The data required are an annual maximum series that is long (25 years is a practical minimum, 30–50 preferable), homogeneous, independent, and stationary — no dam constructed mid-record, no major land-use change, no gauge relocation. The two standard operations are:

1. Assign an empirical recurrence interval by plotting position. Rank the peaks in descending order, $m=1$ for the largest. The Weibull plotting position gives $$T=\frac{n+1}{m}$$ so in a 30-year record the second largest flood on record is assigned $$T=\frac{31}{2}=15.5\ \text{years}$$ Note what this does not mean: it is an estimate of the exceedance frequency of that magnitude, not a statement that the flood recurs every 15.5 years.
2. Fit a distribution and extrapolate beyond the record. An empirical plotting position can never give a 100-year estimate from 30 years of data, so a distribution is fitted — Gumbel (EV1), log-Pearson Type III (the standard for flood frequency in North America) or GEV. Using the Gumbel frequency factor in Chow’s form, $$K_T=-\frac{\sqrt{6}}{\pi}\left[0.5772+\ln\ln\!\left(\frac{T}{T-1}\right)\right]$$ which for $T=100$ years gives $K_{100}=3.137$, so $$x_T=\bar{x}+K_T s = 450 + 3.137\times 120$$ $$\boxed{x_{100}=826\ \text{m}^{3}\text{/s}}$$ The estimate should be reported with confidence limits, which at a 30-year record length are wide — typically of order ±20 % on a 100-year quantile.

Part (iii) — Justification for using probability distributions (4 marks)

Hydrologic processes are driven by atmospheric dynamics that are chaotic on the time scales engineers care about. The magnitude of next year’s flood is not predictable deterministically, but the population of annual floods is statistically stable, and that stability is what makes design possible. Probability distributions are justified on four grounds.

First, they convert an unpredictable variable into a designable one. An engineer cannot know the largest flood in the next 50 years, but can size a structure to a stated exceedance probability — and can state what that means. Second, they permit extrapolation beyond the record: design commonly requires a 100- or 200-year event, while records are 30–60 years long, and only a fitted distribution can bridge that gap defensibly. Third, they make risk explicit and therefore comparable, so that the cost of a larger structure can be weighed against the expected cost of failure, and structures of differing consequence can be assigned differing standards — which is precisely how the Canadian Dam Association ties the inflow design flood to consequence classification. Fourth, they support probabilistic reasoning about combinations of events: the chance of a design storm coinciding with a high tide or an antecedent snowpack.

Example. A culvert is to be designed for a 25-year peak flow. The annual maximum series for the basin is fitted with a log-Pearson III distribution, which returns a 25-year discharge of, say, 96 m³/s. The distribution also states the risk over the structure’s 50-year design life: $1-(1-0.04)^{50} = 87$ % probability of at least one exceedance. That single number changes the design conversation — it tells the owner that the culvert is expected to be exceeded during its life, and that the real design question is what happens when it is (overtopping route, embankment protection), not whether it will be. No deterministic method can produce that statement.

Part (iv) — The meaning of a “100-year flood” (4 marks)

Given. A flood magnitude whose annual exceedance probability is $p = 1/T = 0.01$; a forward window of 25 years.

Find. Whether repeat occurrences are possible, the correct interpretation of the term, and the probability of at least one occurrence in 25 years.

Can two occur within several years, or the same year? Yes — and it is not even rare. The term denotes a magnitude with a 1 % chance of being equalled or exceeded in any given year. Occurrences are conventionally modelled as independent between years, so the fact that one has just happened does not reduce the probability of the next: there is no memory in the system and no schedule. Two events in successive years have a probability of $0.01^2 = 0.0001$ per pair of years, which across the many gauged basins of a province is close to a certainty somewhere each decade. Even two exceedances within a single year are possible, since more than one independent storm can occur in a year; modelling occurrences as a Poisson process with rate $\lambda = 0.01$ per year gives $P(\ge 2 \text{ in one year}) = 1 - e^{-\lambda}(1+\lambda) = 5.0 \times 10^{-5}$, small but not zero.

What is commonly meant. In everyday use the phrase is almost always misunderstood as “a flood that happens once every hundred years”, which invites the fallacy that having had one, a community is safe for a century. The correct meaning is a flood with a 1 % annual exceedance probability — which is why the term 1 % AEP flood is now preferred in Canadian and international practice, and why Canadian floodplain mapping is framed on the 1:100 (1 %) standard rather than on a recurrence interval. The recurrence interval is the mean time between exceedances over a very long period, and the distribution of actual intervals about that mean is extremely wide.

1. Compute the risk of at least one exceedance in 25 years. Successive years are independent Bernoulli trials, so the probability of no exceedance in any single year is $(1-p)$, and over $n$ years the risk of at least one exceedance is $$R = 1-(1-p)^{n} = 1-(1-0.01)^{25}$$ $$\boxed{R = 0.2222 = 22.2\ \%}$$ So there is close to a one-in-four chance that a “100-year flood” occurs within the next 25 years.
2. Extend the same logic to a full 100 years, to expose the fallacy directly. Setting $n = 100$ gives $$R = 1-(0.99)^{100} = 0.634 = 63.4\ \%$$ not 100 %. Over a period equal to the return period there is roughly a 37 % chance the event does not occur at all — the clearest possible demonstration that a return period is not a schedule.