Flood Frequency Analysis

On 6 August 1979, the Machhu-2 dam on the Machhu River in Gujarat, India failed after extreme monsoon rainfall. The wall of water that swept through Morbi killed somewhere between 1,800 and 25,000 people — the death toll was suppressed by the Indian government for decades. Engineers later estimated the peak inflow had been roughly 25 times the design capacity, a flow the dam’s designers had never imagined. The designers had done the right analysis; they had simply worked from records too short to reveal what the river was capable of.

The term “100-year flood” is one of the most misunderstood phrases in civil engineering. It does not mean a flood that occurs once per century. It means a flood with a 1% annual probability of being exceeded — and in any 30-year mortgage period, there is a 26% chance your property experiences at least one such event. Extend that to a bridge designed for a 75-year service life and the exceedance probability climbs above 50%. Engineers need a precise language for rare events: one that connects a probability statement to an actual discharge volume that a spillway, bridge opening, or levee must accommodate.

Flood frequency analysis provides that language. It takes the historical record of annual peak flows — often 40 to 80 years of gauge data — and fits a probability distribution to it. From the fitted distribution, engineers can read off the discharge corresponding to any return period they need: the 10-year flood for a culvert, the 100-year flood for a floodplain map, the 500-year flood for a dam spillway. The method is probabilistic, not deterministic — it cannot predict when the next big flood will come, only how big it is likely to be when it does.

1. The Question

What is the 100-year flood discharge for this river?

Design flood problem:

Every structure near water needs design flood estimate.

Bridge: How high to build?
Dam: What spillway capacity?
Floodplain map: Where is 100-year boundary?
Insurance: What’s the premium?

Flood frequency analysis provides answer:

Statistical method estimating flood magnitude for given return period.

Based on: Historical streamflow records

Annual peak series:

Each year, record maximum instantaneous discharge.

Example - Allegheny River: - 1950: 4,200 cms - 1951: 2,800 cms - 1952: 5,600 cms (major flood) - … - 2025: 3,100 cms

Fit probability distribution to these peaks.

Extrapolate to estimate rare events (100-year, 500-year).

Applications: - Floodplain mapping (FEMA flood insurance rate maps) - Bridge design (scour protection) - Dam spillway sizing - Levee height determination - Building elevation requirements (freeboard) - Emergency management planning

Return period interpretation:

100-year flood: 1% annual exceedance probability (AEP)

Not: “Occurs once per 100 years”

Rather: “1% chance each year, regardless of past floods”

Misconception: “We just had 100-year flood, safe for 100 years”

Wrong! Each year independent.

2. The Conceptual Model

Log-Pearson Type III Distribution

USGS standard method (Bulletin 17C, 2017)

Why logarithms?

Flood peaks positively skewed (long right tail).

Logarithmic transformation normalizes distribution.

Transform to log-space:

Y = \log_{10}(Q)

Pearson Type III distribution:

Three-parameter distribution flexible for skewed data.

f(y) = \frac{\lambda^\alpha}{\Gamma(\alpha)}(y-\beta)^{\alpha-1}e^{-\lambda(y-\beta)}

Where: - \alpha = shape parameter (related to skewness) - \beta = location parameter (related to mean) - \lambda = scale parameter (related to std dev) - \Gamma(\alpha) = gamma function

Method of moments:

Parameters estimated from sample statistics:

Mean: \bar{y} = \frac{1}{n}\sum y_i

Standard deviation: s_y = \sqrt{\frac{1}{n-1}\sum(y_i - \bar{y})^2}

Skew coefficient: G = \frac{n\sum(y_i - \bar{y})^3}{(n-1)(n-2)s_y^3}

Flood quantile formula:

y_T = \bar{y} + K_T \times s_y

Where K_T = frequency factor (function of skew G and return period T)

Back-transform:

Q_T = 10^{y_T}

Frequency Factor Tables

K_T values tabulated by USGS (Bulletin 17C, Appendix 3)

Example values for selected skew and return periods:

Pattern: Higher positive skew → lower K_T for same return period

Interpretation: Positively skewed distributions have heavier upper tail already captured in mean/std dev

Regional Skew

Problem: Sample skew unstable (requires long records)

Solution: Generalized skew from regional analysis

Weighted skew:

G_w = \frac{MSE_G \times G + MSE_r \times G_r}{MSE_G + MSE_r}

Where: - G = station skew - G_r = regional skew (from USGS maps) - MSE = mean square error (weighting factor)

Typical: Regional skew map provides G_r \approx 0 to +0.5 depending on location

3. Building the Mathematical Model

Step-by-Step LP3 Analysis

Step 1: Assemble annual peak data

Minimum 10 years, prefer 30+

Step 2: Check for outliers

High outliers: Peaks unusually large (different flood mechanism?)

Low outliers: Peaks unusually small (dam operation, drought?)

Grubbs-Beck test: Statistical outlier detection

Retain outliers if physically plausible.

Step 3: Transform to logarithms

y_i = \log_{10}(Q_i)

Step 4: Calculate sample statistics

\bar{y} = \frac{1}{n}\sum_{i=1}^n y_i

s_y = \sqrt{\frac{1}{n-1}\sum_{i=1}^n (y_i - \bar{y})^2}

G = \frac{n}{(n-1)(n-2)s_y^3}\sum_{i=1}^n (y_i - \bar{y})^3

Step 5: Apply weighted skew (combine station and regional)

Step 6: Select frequency factors from tables

Step 7: Calculate flood quantiles

y_T = \bar{y} + K_T(G_w, T) \times s_y

Q_T = 10^{y_T}

Step 8: Estimate confidence intervals

SE(y_T) = s_y \sqrt{\frac{1 + K_T^2/2}{n}}

95% CI: y_T \pm 1.96 \times SE(y_T)

Plotting Position

Empirical probability for observed peaks:

P_i = \frac{i}{n+1}

Where i = rank (1 = largest)

Return period:

T_i = \frac{1}{P_i}

Plot observed (T_i, Q_i) vs fitted curve Q_T

Visual check of distribution fit

4. Worked Example by Hand

Problem: Estimate design floods for bridge design.

River: Tributary with 15 years of annual peak data

Annual peaks (m³/s):

Regional skew: G_r = +0.2 (from USGS map)

Calculate 10-year, 50-year, and 100-year floods.

Solution

Step 1: Transform to logarithms

y_i = \log_{10}(Q_i)

Step 2: Calculate mean

\bar{y} = \frac{1}{15}(2.653 + 2.763 + ... + 2.724) = \frac{40.897}{15} = 2.726

Step 3: Calculate standard deviation

s_y = \sqrt{\frac{1}{14}\sum(y_i - 2.726)^2}

Example deviation: (2.653 - 2.726)^2 = 0.00533

Sum of squared deviations = 0.2156

s_y = \sqrt{\frac{0.2156}{14}} = \sqrt{0.0154} = 0.124

Step 4: Calculate station skew

G = \frac{15}{14 \times 13 \times 0.124^3}\sum(y_i - 2.726)^3

Example cubed deviation: (2.653 - 2.726)^3 = -0.000389

Sum of cubed deviations = +0.000821

G = \frac{15 \times 0.000821}{182 \times 0.00191} = \frac{0.0123}{0.348} = +0.035

Low skew (nearly symmetric)

Step 5: Weighted skew

Assume MSE_G = 0.30 (typical for n=15), MSE_r = 0.25

G_w = \frac{0.30 \times 0.035 + 0.25 \times 0.2}{0.30 + 0.25} = \frac{0.0105 + 0.050}{0.55} = 0.110

Use G_w = 0.11 for frequency factors

Step 6: Frequency factors (interpolated from tables)

For G = 0.11: - K_{10} = 1.27 - K_{50} = 2.01 - K_{100} = 2.29

Step 7: Calculate flood quantiles

10-year flood:

y_{10} = 2.726 + 1.27 \times 0.124 = 2.726 + 0.157 = 2.883

Q_{10} = 10^{2.883} = 764 \text{ m}^3\text{/s}

50-year flood:

y_{50} = 2.726 + 2.01 \times 0.124 = 2.726 + 0.249 = 2.975

Q_{50} = 10^{2.975} = 944 \text{ m}^3\text{/s}

100-year flood:

y_{100} = 2.726 + 2.29 \times 0.124 = 2.726 + 0.284 = 3.010

Q_{100} = 10^{3.010} = 1023 \text{ m}^3\text{/s}

Step 8: Confidence intervals (95%)

SE(y_{100}) = 0.124 \sqrt{\frac{1 + 2.29^2/2}{15}} = 0.124 \sqrt{\frac{3.62}{15}} = 0.124 \times 0.491 = 0.061

y_{100} \pm 1.96 \times 0.061 = 3.010 \pm 0.120

Range: y = 2.890 to $3.130$

Q_{100} = 10^{2.890} \text{ to } 10^{3.130} = 776 \text{ to } 1349 \text{ m}^3\text{/s}

Wide uncertainty: 776-1349 cms (±28%)

Step 9: Design recommendation

Bridge design: Use upper confidence limit for safety

Design Q₁₀₀ = 1350 m³/s

5. Computational Implementation

Below is interactive flood frequency analyzer.

<label>
  Record length (years):
  <input type="range" id="record-length" min="10" max="100" step="5" value="30">
  <span id="length-val">30</span>
</label>
<label>
  Skew coefficient:
  <input type="range" id="skew" min="-0.5" max="1.5" step="0.1" value="0.2">
  <span id="skew-val">0.2</span>
</label>
<label>
  Mean flow (log cms):
  <input type="range" id="mean-log" min="2.0" max="3.5" step="0.1" value="2.7">
  <span id="mean-val">2.7</span>
</label>
<div class="flood-info">
  <p><strong>Q₁₀:</strong> <span id="q10">--</span> m³/s</p>
  <p><strong>Q₅₀:</strong> <span id="q50">--</span> m³/s</p>
  <p><strong>Q₁₀₀:</strong> <span id="q100">--</span> m³/s</p>
  <p><strong>CI width (Q₁₀₀):</strong> ±<span id="ci-width">--</span>%</p>
</div>

6. Summary

Flood frequency analysis estimates design floods via Log-Pearson Type III distribution fitted to annual peak streamflow data transforming to logarithmic space for skewed distributions. Frequency factors K_T derived from skew coefficient and return period enabling quantile calculation via y_T = mean + K_T × std_dev. USGS Bulletin 17C standard method incorporating regional skew weighting and outlier detection improving estimates. Confidence intervals widen significantly for rare events with ±20-40% typical for 100-year estimates from 30-year records. Applications span bridge design, dam spillways, floodplain mapping, and flood insurance requiring return periods from 10 to 10,000 years depending on consequence of failure.