Muhammad Nura¹, Zahrahtul Amani Zakaria^2,

Terengganu, on the northeast coast of Peninsular Malaysia, is a flood-prone region whose annual maximum daily rainfall records are strongly positively skewed owing to the Northeast Monsoon (NEM). This skewness can bias conventional regional frequency analysis (RFA) based on L-moments. Trimmed L-moments (TL-moments) provide a more robust alternative by down-weighting extreme order statistics. Although L-moment RFA is well established for Malaysian rainfall, no study has combined Ward's hierarchical clustering with both L-moments and TL-moments at trimming levels t = 1 and t = 2 within Terengganu, nor systematically evaluated four candidate distributions GEV, GLO, GPA, and the three-parameter Kappa Type-II (K3D-II) under all estimation methods in parallel. Annual maximum daily rainfall records from nine Department of Irrigation and Drainage (DID) stations (1971–2023; 252 station-years) were analysed. Ward's minimum-variance clustering of normalised TL-moment ratio site characteristics, validated by the average silhouette width (ASW) criterion, identified two homogeneous regions. Discordancy screening, heterogeneity testing via 500 Monte Carlo simulations, and Z-statistic goodness-of-fit tests were applied to all four distributions under each estimation method. Normalised growth-factor quantiles were estimated at return periods T = 2–200 years. Ward's clustering yielded an optimal two-region solution (k = 2; ASW = 0.6472): Region R1 (seven stations, low-to-moderate skewness) and Region R2 (two stations most affected by the December 2013 NEM event). Both regions were acceptably homogeneous (H1 < 1) under all methods. TL-moments progressively attenuated regional skewness relative to L-moments, with attenuation most pronounced in R2 (up to 59.7% at t = 2), reflecting reduced influence of the 2013 extreme observations on parameter estimation. The best-fit distribution varied by both region and estimation method: K3D-II performed best for R1 under L-moments and for R2 under TL-moments, while GLO and GEV were preferred for R1 at t = 1 and t = 2, respectively. TL-moment growth factors were 7–59% lower than their L-moment counterparts for T ≥ 20 years. These results suggest that TL-moments combined with the K3D-II distribution provide a more robust framework for regional rainfall frequency analysis in high-skewness, NEM-dominated environments. L-moment and TL-moment estimates are recommended for joint use as upper and lower design bounds, with the appropriate choice guided by the consequence class of the hydraulic structure.

homogeneous region delineation, moment-ratio diagram, hydro-climatic sub-regions, quantile attenuation, design rainfall bounds, parameter estimation robustness, outlier-robust estimation