Reliable estimates of flood quantiles are a prerequisite for the safe design of hydraulic structures and for flood risk management. Traditionally, they are derived from annual maximum series that retain one event per year, discarding secondary peaks and small maxima from dry years. This study estimated the frequency of maximum discharges of the Erzen River at Ndroq station using a partial duration series built with the peaks-over-threshold approach, and compared its results with the conventional annual maximum series analysis. The daily discharge record (1979–2023) was screened over 100 candidate thresholds, and the threshold was selected by reading six diagnostic plots. Three are classical: the number of exceedances, the Generalised Pareto shape parameter and the mean excess plot. Three are rarely combined: the threshold-wise Anderson-Darling test of the Generalised Pareto fit, its parametric-bootstrap p-value and the dispersion index test of the Poisson occurrence model. Independence of the retained peaks was enforced by the Bulletin 17B inter-event time criterion (11 days) together with the Flood Estimation Handbook recession criterion (discharge receding below two-thirds of the preceding peak), and the sensitivity of the results to the time separation window was quantified. At the selected threshold of 445 m³/s the series comprises 34 independent events (0.76 events per year) whose magnitudes follow a Generalised Pareto distribution (shape κ = −0.276) fitted by L-moments, while the annual event counts are consistent with a Poisson process (dispersion index 0.91). Annual exceedance probabilities were obtained by the empirical Langbein transformation and by the analytical conversion of the fitted Pareto–Poisson model into an equivalent Generalised Extreme Value distribution. The 90% confidence intervals were computed by parametric bootstrap. The partial duration series yields a 1% annual exceedance probability discharge of 1114 m³/s, approximately 12% higher than the 996 m³/s obtained from the annual maximum series. The difference grows to about 51% at the 1000-year return period. Half of the annual maximum series members lie below the threshold, which produces a near-Gumbel upper tail that the threshold-censored sample contradicts. The peaks-over-threshold approach provides a more representative description of flood behaviour in the Erzen River and a more conservative basis for the design of hydraulic works and flood mitigation planning.