a comparative study of imputation techniques for hydrometeorological data of ravand river in karkhe basin

Reliable streamflow records are essential for flood forecasting, reservoir management, and hydrological modeling, yet data discontinuities are widespread due to sensor faults, telemetry dropouts, and operational limitations. this study systematically evaluates a broad range of imputation methods for daily streamflow reconstruction at the badrgerd station in kermanshah, iran (2009–2021), using meteorological predictors (temperature and precipitation) as auxiliary variables. a total of 13 approaches were benchmarked, spanning deep learning (hl-vae, vae, lstm, transformer, tcn, autoencoder, gain), tree-based ensembles (xgboost, gradient boosting, missforest), and classical statistical methods (em, mice, gaussian process, linear regression). performance was assessed on a withheld test set comprising 10% of the dataset using three efficiency coefficients: mean absolute error (mae), coefficient of determination (r²), and percent bias (pbias). results demonstrate that the hl-vae achieved the most balanced performance (lowest mae, highest r², and near-zero bias), confirming its strength in capturing nonlinear dependencies. lstm also provided competitive results, while em maintained robust yet less explanatory accuracy. tree-based methods showed larger errors and negative bias, and the autoencoder and tcn recorded the weakest outcomes. these findings highlight the superior capability of deep generative models—particularly hl-vae—for imputing missing hydrometeorological data.