Adaptive Time-Lagged Ensemble for Short-Range Streamflow Prediction Using WRF-Hydro and LDAPS

Abstract

This study evaluates a time-lagged ensemble averaging strategy to improve the accuracy and robustness of short-range streamflow point forecasts when hydrological simulations are driven by deterministic numerical weather prediction (NWP) forcing. We implemented WRF-Hydro in standalone mode for the Geumho River basin, South Korea, using Local Data Assimilation and Prediction System (LDAPS) forecasts initialized every 6 h with lead times up to 48 h. Time-lagged ensembles were constructed by averaging overlapping WRF-Hydro predictions from successive LDAPS initializations. Across two contrasting flood-producing storms, ensemble-mean forecasts consistently reduced lead-time-dependent skill degradation relative to single-initialization forecasts; the event-wise median Nash–Sutcliffe efficiency at the downstream gauge improved from 0.39 to 0.81 at 48 h (Event 2020) and from 0.48 to 0.85 at 24 h (Event 2022), while RMSE decreased by up to 48%. The most effective ensemble window varied with storm evolution and forecast horizon, indicating additional gains from adaptive time-lag selection. Overall, time-lagged ensemble averaging provides a practical, low-cost post-processing approach to enhance operational short-range streamflow prediction with NWP forcings.

1. Introduction

Accurate short-range streamflow forecasting remains a persistent challenge in operational hydrology due to its critical role in flood preparedness, water resource planning, and emergency response decisions [1,2]. Forecast accuracy is fundamentally constrained by uncertainties in meteorological inputs, hydrological model structures, and parameterization schemes [3,4,5]. These limitations become especially pronounced during extreme events, when deterministic forecasts often fail to capture the full range of possible outcomes, thereby reducing the timeliness and effectiveness of risk management efforts [6,7].

Advances in high-resolution, short-range numerical weather prediction (NWP) models have significantly improved the quality of meteorological forcing available for hydrological applications [8,9]. When used to drive physically based, distributed hydrological models, these forecasts enable more detailed simulation of watershed responses. However, the inherent uncertainty in meteorological forecasts often propagates through the hydrological modeling chain, resulting in considerable errors in streamflow predictions, especially during high-impact events [10,11,12].

WRF-Hydro is a fully distributed, physically based hydrological model that tightly integrates atmospheric and land surface processes [13]. The model has demonstrated robust capabilities in simulating rainfall-runoff processes, channel routing, and inundation dynamics under a wide range of meteorological conditions [14,15,16,17]. It also forms the basis of the U.S. National Water Model (NWM), which employs a seven-member time-lagged ensemble system using successively initialized Global Forecast System (GFS) forecasts to extend predictive lead time up to 10 days [18]. This ensemble approach has been successfully implemented at the continental scale.

Ensemble forecasting techniques, particularly time-lagged ensembles, have attracted increasing attention as effective approaches to quantify and reduce forecast uncertainty without necessitating physically perturbed model inputs [19,20,21,22]. By overlapping forecasts initialized at successive times, time-lagged ensembles create multiple realizations of hydrological predictions, facilitating uncertainty assessment and enhancing forecast reliability through ensemble averaging [23,24,25]. Concurrently, multi-model ensembles have also been effective in mitigating structural biases and improving prediction accuracy by integrating diverse modeling structures [26,27].

Despite the proven advantages of ensemble forecasting, most operational implementations employ fixed time-lagged ensemble configurations regardless of forecast horizon or meteorological conditions [28,29,30]. While computationally efficient, such fixed approaches may not fully exploit the varying predictability characteristics associated with different lead times and event types. This study addresses this gap by systematically evaluating how optimal ensemble window lengths vary with forecast lead times and meteorological event characteristics. We demonstrate that, rather than relying on a single fixed configuration, adapting ensemble windows to specific event types can significantly improve prediction accuracy. This adaptive strategy is evaluated using high-resolution distributed modeling (100 m) to effectively capture spatially varying watershed responses.

This study assesses the effectiveness of adaptive time-lagged ensemble forecasts in improving short-range streamflow predictions (up to 48 h) using the distributed hydrological model WRF-Hydro driven by Local Data Assimilation and Prediction System (LDAPS) meteorological forecasts. The Geumho River basin in South Korea, known for its distinct hydrological responses during monsoon and typhoon events, is selected to evaluate ensemble forecasting across two contrasting scenarios: prolonged intense rainfall (Event 2020) and a brief, high-intensity typhoon (Event 2022). Specifically, the research investigates: (1) whether time-lagged ensemble predictions outperform deterministic forecasts; (2) how effectively adaptive ensemble strategies enhance predictive accuracy and reduce uncertainty; and (3) how optimal ensemble configurations vary with the meteorological characteristics of each event. The outcomes provide critical insights for tailored ensemble prediction strategies to enhance flood management and risk mitigation in monsoon-prone regions.

2. Materials and Methods

2.1. Hydrological Model

WRF-Hydro, a key component of the NWM, is an integrated hydrological modeling framework developed by the National Center for Atmospheric Research (NCAR). The model supports two operational modes: a coupled mode, which integrates directly with the WRF atmospheric model, and a standalone mode, which operates using external meteorological inputs [13].

In this study, high-resolution short-term streamflow prediction was performed using WRF-Hydro version 5.2, specifically configured in standalone mode with meteorological forecasts provided externally from LDAPS. The WRF-Hydro modeling system has a modular structure, where selected physical modules are linked via the WRF-Hydro Driver/Coupler. Each module can be configured at different spatial resolutions, enabling high-resolution hydrological simulations. Surface runoff, groundwater flow, and channel routing modules were utilized in the simulations, and the detailed workflow illustrating how LDAPS meteorological forecasts were provided as external inputs to drive WRF-Hydro simulations in standalone mode is presented in Figure 1.

The land surface model used the Noah-Multiparameterization (Noah-MP) Land Surface Model, which has strengths in runoff analysis, to simulate vertical hydrological processes such as evapotranspiration, soil moisture, and energy flux. Noah-MP LSM extends the surface simulation process of the existing WRF model to include detailed hydrological processes, including surface runoff, channel routing, and groundwater flow. The soil layer consists of four layers, reflecting the soil characteristics of the surface layer (0–10 cm) and deep layer (10–40 cm) to simulate vertical water movement. In this study, four key Noah-MP land-surface and groundwater-related parameters were calibrated: $Z_{\mathrm{m}\mathrm{a}\mathrm{x}}$ (maximum soil moisture capacity), Coeff (groundwater bucket coefficient), Expon (groundwater bucket exponent), and REFKDT (surface runoff parameter). Parameter descriptions and the default and calibrated values are provided in Supplementary Table S1.

The surface tracking model simulates surface flow and subsurface flow. Surface water flow is calculated based on the continuity Equation (1) and Manning’s Equation (2).

$${\displaystyle \frac{\partial h}{\partial t}}+{\displaystyle \frac{\partial q_x}{\partial x}}+{\displaystyle \frac{\partial q_y}{\partial y}}={\mathcal{i}}_e$$

(1)

$$q_x={\displaystyle \frac{\sqrt{S_{fx}}}{n_{\mathrm{O}\mathrm{V}}}}{h}^{5/3}$$

(2)

where $h$ is water depth (m), $q_x$ and $q_y$ are unit discharge in $x$ and $y$ directions (${\mathrm{m}}^2/\mathrm{s}$), $t$ is time, ${\mathcal{i}}_e$ is infiltration excess runoff ($\mathrm{m}/\mathrm{s}$), $n_{\mathrm{O}\mathrm{V}}$ is Manning’s roughness coefficient ($s/{\mathrm{m}}^{1/3}$), $S_{fx}$ is friction slope.

Channel routing applied a grid-based routing algorithm of the diffusive wave equation. The continuity equation in the diffusive wave method is Equation (3) and the momentum Equation (4).

$${\displaystyle \frac{\partial A}{\partial t}}+{\displaystyle \frac{\partial Q}{\partial x}}=q_{lat}$$

(3)

$${\displaystyle \frac{\partial Q}{\partial t}}+{\displaystyle \frac{\partial }{\partial x}}\left({\beta }_m{\displaystyle \frac{Q^2}{A}}\right)+gA{\displaystyle \frac{\partial Z}{\partial x}}=-gA{S}_f$$

(4)

where $x$ is channel flow direction (m), $A$ is flow cross-sectional area (${\mathrm{m}}^2$), $Q$ is discharge (${\mathrm{m}}^3/\mathrm{s}$), $q_{lat}$ is lateral inflow to the channel grid (${\mathrm{m}}^2/\mathrm{s}$), ${\beta }_m$ is momentum correction factor, $Z$ is water level (m), $g$ is gravitational acceleration ($9.81 \mathrm{m}/{\mathrm{s}}^2$), and $S_f$ is friction slope.

2.2. Study Area and Data

The study area is the Geumho River catchment, a tributary of the Nakdong River, with a river length of 114.6 km and a catchment area of 2087.9 km², accounting for approximately 9.2% of the entire Nakdong River basin (Figure 2). Located in the middle reaches of the Nakdong River, the Geumho River originates from Pohang, and flows southward through Yeongcheon and Gyeongsan cities. The catchment has an annual mean temperature of 13 °C and annual precipitation of 1007 mm [31].

2.2.1. Topographic and Land-Use Data

The WRF-Hydro model requires geographic information such as land-use data, soil maps, and Digital Elevation Model (DEM) as main topographic input data (Figure 2), which are converted into model input data through the WRF Preprocessing System (WPS). DEM data serve as the basic input data for high-resolution terrain routing and was constructed by resampling the 30 m spatial resolution DEM data provided by NASA Shuttle Radar Topography Mission (SRTM) to 100 m resolution.

Land-use data are crucial input information representing surface characteristics in WRF-Hydro and were constructed using the medium-class land-cover map provided by the Ministry of Environment (Figure 3a). As shown in Figure 3a, deciduous broadleaf forest shows the highest distribution at approximately 28.9%, followed by evergreen needle leaf forest at about 24.8%, and urban and built-up land at about 9.9%.

Soil data were obtained from the Rural Development Administration’s 1:25,000 detailed soil map, which classifies soils into a surface layer (0–10 cm) and a subsoil layer (10–40 cm). Figure 3b,c shows the detailed soil maps of surface and deep layers in the Geumho River catchment. The surface-layer soil (Figure 3b) is dominated by silt loam at about 46.8%, followed by loam at about 28.8%, and sandy loam at about 15.8%. For the deep layer (Figure 3c), clay loam accounts for the highest proportion at about 57.6%, followed by sandy loam at about 16.5%, and silty clay loam at about 10.9%.

2.2.2. Meteorological Data

The WRF-Hydro simulation requires eight meteorological elements (

Table 1. Mapping of LDAPS variables to WRF-Hydro forcing elements.

WRF-Hydro Variables	LDAPS Variables	Description	Unit
SWDOWN	TDSW	Incoming shortwave radiation	$\mathrm{W}/{\mathrm{m}}^2$
LWDOWN	DLSW	Incoming longwave radiation	$\mathrm{W}/{\mathrm{m}}^2$
Q2D	SPFH	Specific humidity	$\mathrm{k}\mathrm{g}/\mathrm{k}\mathrm{g}$
T2D	TMPR	Air temperature	$\mathrm{K}$
PSFC	PRES	Surface pressure	$\mathrm{P}\mathrm{a}$
U2D	UGRD	Near surface wind in the u-component	$\mathrm{m}/\mathrm{s}$
V2D	VGRD	Near surface wind in the v-component	$\mathrm{m}/\mathrm{s}$
RAINRATE	LSPR	Precipitation rate	$\mathrm{m}\mathrm{m}/\mathrm{s}$ or $\mathrm{k}\mathrm{g}/{\mathrm{m}}^2/\mathrm{s}$

): incoming shortwave radiation (SWDOWN), incoming longwave radiation (LWDOWN), specific humidity (Q2D), air temperature (T2D), surface pressure (PSFC), east–west wind (U2D), north–south wind (V2D), and precipitation rate (RAINRATE), utilizing both numerical weather prediction data and ground observation data.

The LDAPS produces hourly weather forecast information up to 48 h for the Korean Peninsula region. LDAPS is executed four times daily (00, 06, 12, and 18 UTC), each providing up to 48 h forecasts. It uses 3 h interval forecast outputs from the global model GDAPS as lateral boundary conditions. Based on the Unified Model (UM) developed in the UK, it has a horizontal spatial resolution of 1.5 km and consists of 70 vertical layers up to approximately 40 km.

LDAPS numerical simulation results are provided through the Korea Meteorological Data Portal in three types: isobaric surface, single level, and model level, in GRIB2 format. The necessary elements for WRF-Hydro meteorological forcing, as listed in Table 1, were extracted from LDAPS data and converted to netCDF format to construct the hourly input data. LDAPS fields at 1.5 km resolution were spatially interpolated to the 100 m WRF-Hydro grid using bilinear interpolation, maintaining the original hourly temporal resolution across all meteorological variables listed in Table 1. For example, precipitation information was applied as WRF-Hydro input data by extracting the Large-Scale Precipitation Rate (LSPR) variable from LDAPS forecast data.

Additionally, point observation weather data provided by the Korea Meteorological Administration were utilized for the verification of the WRF-Hydro model. The observational data set consisted of a total of 11 stations within the Geumho River basin, including nine Automatic Weather System (AWS) and two Automated Synoptic Observing System (ASOS) stations. Eight meteorological elements (temperature, precipitation, wind speed, wind direction, atmospheric pressure, humidity, sunshine duration, and solar radiation) measured at these stations were compiled at 1 h intervals as input data for the WRF-Hydro model.

The simulation periods were selected to cover two significant hydrological events. The first event (Event 2020) covers the August 2020 heavy rainfall period from 00 UTC August 3 to 00 UTC August 14, capturing a series of intensive rainfall events in the central region of South Korea. The second event (Event 2022) spans from 00 UTC September 1 to 00 UTC September 12, encompassing the passage of Typhoon Hinnamnor. These periods were chosen to ensure sufficient coverage of both the rainfall events and their resulting hydrological responses in the catchment.

2.3. Ensemble Forecast Configuration

2.3.1. Time-Lagged Ensemble Methodology

To quantify and reduce the uncertainty inherent in streamflow forecasts derived from deterministic LDAPS inputs, this study implemented a time-lagged ensemble approach at the output level of the WRF-Hydro model output. Time-lagged ensembles represent a cost-effective method for generating probabilistic forecasts by utilizing forecast outputs from multiple initialization times that share common valid forecast periods [20].

The fundamental principle of the time-lagged ensemble is to leverage the temporal overlap of forecasts initialized at different times. LDAPS forecasts are operationally issued four times daily at 6 h intervals (00:00, 06:00, 12:00, and 18:00 UTC), with each forecast providing meteorological predictions extending up to 48 h into the future. These forecasts serve as meteorological forcing inputs to WRF-Hydro simulations, which generate corresponding 48 h streamflow predictions. Figure 4 provides an example illustration of time-lagged rainfall forecasts and streamflow simulations valid at 00 UTC 6 September from eight ensemble members initialized at different times, ranging from 00 UTC 4 September to 18 UTC 5 September. This visualization demonstrates how forecasts from multiple initialization times can target the same valid forecast hour, thereby providing multiple realizations of the same meteorological event. This approach exploits the fact that forecasts initialized at different times but targeting the same valid forecast hour can be combined to form an ensemble.

2.3.2. Ensemble Configuration Across Lead Time

Figure 5 illustrates the configuration of the time-lagged ensemble system implemented in this study, where ensembles are constructed by selecting different portions of each WRF-Hydro simulation to evaluate forecast performance across varying lead times. The number of ensemble members systematically varies depending on the forecast lead time, with shorter lead times containing fewer ensemble members due to fewer overlapping initializations. Specifically, ensemble predictions at the 12 h lead time interval (Ens-12h) consist of two ensemble members, those at the 24 h interval (Ens-24h) consist of four members, the 36 h interval (Ens-36h) consists of six members, and the 48 h interval (Ens-48h) comprises the maximum of eight members. This configuration reflects the availability of overlapping forecasts, where longer lead times can utilize more initialization times that contribute to the same forecast period.

For performance evaluation, two types of predictions are compared: (1) Ensemble Mean—the average of ensemble members from initialization times before the current forecast time, and (2) Single—an individual forecast member from a single initialization time that represents a deterministic prediction without ensemble averaging.

2.4. Prediction Evaluation Metrics

In this study, hydrological observation data within the Geumho River basin was utilized to verify the flow simulation results of the WRF-Hydro model. Given the locations and data availability within the basin, Ansim Bridge and Kangchang Bridge were selected as the main discharge comparison points. Ansim Bridge, located in the middle reaches of the Geumho River, reflects the runoff characteristics after the confluence of major upstream tributaries, while Kangchang Bridge, situated at the lowermost part of the basin, represents the runoff of the entire Geumho River basin.

The performance of streamflow predictions was evaluated using both deterministic and probabilistic verification metrics. For deterministic evaluation, Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), and Root Mean Square Error (RMSE) were used.

NSE is a dimensionless metric evaluating the agreement between observation and simulation, ranging from $\infty$ to 1, with 1 indicating perfect agreement:

$$NSE=1-{\displaystyle \frac{{\sum }_{t=1}^N{(Q_t^{obs}-Q_t^{sim})}^2}{{\sum }_{t=1}^N{(Q_t^{obs}-Q_{mean}^{sim})}^2}}$$

(5)

where $Q_t^{obs}$ is the observation at time $t$, $Q_t^{sim}$ is the simulation at time t, and $Q_{mean}^{sim}$ is the mean of observations.

KGE considers correlation, variability bias, and mean bias comprehensively:

$$KGE=1-\sqrt{{(\gamma -1)}^2+{(\alpha -1)}^2+{(\beta -1)}^2}$$

(6)

where $\gamma$ is the correlation coefficient between simulation and observation, $\alpha$ represents the ratio of simulated and observed variability, and $\beta$ is the ratio of simulated and observed means.

RMSE quantifies the average magnitude of prediction errors in the same unit as the variable being predicted:

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\sum }_{t=1}^N{(Q_t^{obs}-Q_t^{sim})}^2}$$

(7)

where $Q_t^{sim}$ is the simulated discharge at time t, $Q_t^{obs}$ is the observed discharge at time t, and $N$ is the total number of time steps.

Percent Bias (PBIAS) was also used to evaluate the model’s performance. PBIAS measures the average tendency of the simulated values to be larger or smaller than their observed counterparts, expressed as a percentage. The optimal value is 0, with positive values indicating model overestimation bias and negative values indicating underestimation bias:

$$PBIAS={\displaystyle \frac{{\sum }_{t=1}^N{Q}_t^{sim}-Q_t^{obs}}{{\sum }_{t=1}^N{Q}_t^{obs}}}\times 100$$

(8)

where $Q_t^{sim}$ is the simulated discharge at time t, $Q_t^{obs}$ is the observed discharge at time t, and $N$ is the total number of time steps.

3. Results

3.1. Comparison of Streamflow Prediction Time Series

This subsection visually compares deterministic (single forecast) and time-lagged ensemble mean forecasts in predicting streamflow using the WRF-Hydro model driven by LDAPS meteorological inputs. The adaptive approach evaluated here refers to varying ensemble configurations (Ens-12h through Ens-48h) based on forecast lead time and event characteristics, where optimal configurations are identified retrospectively based on performance evaluation. The comparison examines two hydrological extremes with contrasting characteristics, prolonged intense rainfall (Event 2020) and a rapidly evolving typhoon (Event 2022), highlighting both the merits of ensemble forecasts over single forecasts and the necessity of adaptive ensemble configurations.

Figure 6 presents observed and predicted streamflow at the Ansim and Kangchang stations and precipitation during Event 2020 (3–14 August 2020) and Event 2022 (Typhoon Hinnamnor, 1–12 September 2022). Single forecasts (individual ensemble members, red lines) and ensemble mean forecasts (ensemble range, blue shading) are compared against observed streamflow (black circles) and ground-observation-driven simulations (black dashed line). Detailed precipitation forecast analysis is presented in Section 3.2.1. The validation periods shown in Figure 6 specifically cover intervals when streamflow sharply increased due to extreme rainfall, and these periods will be used for detailed analysis in later sections assessing adaptive ensemble configurations.

Event 2020 exhibited prolonged flooding characterized by gradual rises and recessions, with peak streamflows around 1800 m³/s at Ansim and 2500 m³/s at Kangchang. In contrast, Event 2022, influenced by intense rainfall from Typhoon Hinnamnor, displayed sharp peaks and rapid recessions, reaching approximately 1500 m³/s at Ansim and 1300 m³/s at Kangchang. These distinct responses underscore the importance of forecasting strategies adaptable to varied meteorological conditions.

Visual inspection demonstrates that the 48 h ensemble mean forecasts consistently offered improved reliability and reduced prediction uncertainties compared to single deterministic forecasts. Nonetheless, the ensemble mean occasionally underestimated peak flows, highlighting limitations in capturing extreme hydrological events. Conversely, single forecasts occasionally provided better peak flow predictions due to their broader spread, though accompanied by greater uncertainty and reduced overall reliability. These findings emphasize the operational value of ensemble forecasts while also suggesting the need for careful optimization of ensemble configurations. The subsequent sections will further quantitatively analyze precipitation ensemble characteristics (Section 3.2.1) and these validation periods under various ensemble configurations (Section 3.2.2 and Section 3.2.3) to identify optimal forecasting strategies.

3.2. Analysis of Ensemble Forecast Performance by Lead Time

This subsection evaluates how effectively time-lagged ensemble forecasts predicted hydrological responses across various lead times. The analysis begins with an assessment of rainfall predictions from LDAPS against ground observations (Section 3.2.1). Subsequently, adaptive ensemble prediction performance is analyzed to determine the optimal ensemble configurations across different forecast horizons (Section 3.2.2). Finally, a statistical evaluation of the ensemble forecasts is provided, focusing on the improvements over single deterministic forecasts with varying lead-time windows (Section 3.2.3).

3.2.1. Rainfall Prediction Analysis

Figure 7 compares temporally averaged LDAPS rainfall predictions with ground observations from AWS and ASOS stations for the two selected hydrological events. The analysis indicates notable differences in rainfall patterns between the two events, with LDAPS generally exhibiting stable performance in timing but systematic underestimation of rainfall amounts, particularly at longer lead times.

During Event 2020 (Figure 7a,b), a prolonged heavy rainfall scenario was observed, characterized by multiple moderate-to-high-intensity precipitation peaks spanning from August 3 to 14, with a total of 121 precipitation hours and maximum precipitation intensity of 14.6 mm/h among all ensemble members (

Table 2. Precipitation event characteristics and ensemble spread statistics for the Ens-48h configuration. Mean and maximum ensemble spread indicate temporal average and peak standard deviation across ensemble members.

Event	Maximum Precipitation Intensity (mm/h)	Total Precipitation Hours (h)	Mean Ensemble Spread (mm/h)	Maximum Ensemble Spread (mm/h)
Event 2020	14.6	121	0.9	9.1
Event 2022	18.2	51	0.3	6.4

). As shown in Figure 7a,b, LDAPS demonstrated good temporal alignment with observed rainfall peak, but consistently underestimated rainfall magnitudes across varying lead times. The ensemble exhibited mean ensemble spread of 0.9 mm/h with gradually increasing variability over time (Figure 6a), reflecting sustained forecast uncertainty throughout the multi day event.

Event 2022 (Figure 7c,d), associated with Typhoon Hinnamnor, exhibited a markedly different rainfall pattern. Rainfall during the typhoon displayed maximum intensity of 18.2 mm/h among ensemble members, concentrated within a shorter total of 51 precipitation hours (Table 2). The ensemble showed mean ensemble spread of 0.3 mm/h with concentrated variability during peak rainfall periods (Figure 6b), suggesting more focused forecast uncertainty during the typhoon passage. This scenario was characterized by a rapid onset and equally rapid cessation of rainfall. Although LDAPS forecasts successfully captured the timing of this concentrated rainfall event, they showed more pronounced underestimation of peak intensities compared to Event 2020. The intensity and rapid variability in the spatial distribution of this rainfall event created greater challenges for deterministic forecasts, reinforcing the value of time-lagged ensemble approaches in stabilizing prediction performance under conditions of extreme spatial and temporal rainfall variability.

Single deterministic forecasts exhibit substantial uncertainty, as some rainfall and streamflow examples are shown in Figure 4 and Figure 6, occasionally resulting in significantly inaccurate predictions, especially during periods of high rainfall intensity and spatial variability. The time-lagged rainfall ensemble approach reduces these uncertainties by averaging predictions across multiple initializations, thereby providing more stable and accurate rainfall estimates. This attenuation of uncertainties in rainfall predictions subsequently contributes to the enhanced performance observed in time-lagged ensemble streamflow forecasts compared to single deterministic forecasts. The effectiveness of adaptive ensemble strategies, tailored to specific hydrological event characteristics, is further assessed in the following subsection.

3.2.2. Adaptive Ensemble Prediction Performance Analysis

Performance evaluation was conducted by analyzing forecast accuracy across different time-lagged ensemble configurations. Each configuration (Ens-12h, Ens-24h, Ens-36h, Ens-48h) was evaluated by comparing ensemble mean predictions with observations over the respective forecast periods.

For Event 2020, the Ens-12h ensemble and the single prediction produced similar results at short lead times (up to 12 h), with NSE values above 0.85 and RMSE ranging between 200 and 250 m³/s. At the 24 h lead time, however, prediction accuracy declined in the Ens-24h configuration, with NSE decreasing to 0.68 and RMSE increasing to around 320 m³/s. This temporary decrease in performance at Kangchang appears to result from the inclusion of single predictions at the end of the 24 h ensemble window. Due to the time-lagged structure, the ensemble mean includes forecasts made at later initialization times, during which prediction accuracy was lower. These lower-performing forecasts were averaged into the ensemble, reducing overall skill near the tail end of the lead time (Figure 8).

At the 36 h and 48 h lead times, ensemble configurations outperformed single predictions. Ens-36h and Ens-48h maintained NSE values above 0.8 and reduced RMSE to 260–280 m³/s. In contrast, single predictions showed a sharp decline in accuracy, with NSE falling below 0.5 and RMSE exceeding 600 m³/s at Kangchang. PBIAS also improved in longer ensemble configurations, with values approaching zero, indicating effective bias mitigation (Figure 8).

For Event 2022, prediction patterns at Kangchang differed significantly. Single predictions consistently underperformed across all lead times, with NSE between 0.3 and 0.5 and RMSE around 250–300 m³/s. PBIAS values indicated a persistent underestimation bias of approximately 40% (Figure 9). Among the ensemble configurations, Ens-24h produced the best performance at the 24 h lead time, with NSE increasing to 0.85 and RMSE decreasing to 141 m³/s. PBIAS also improved slightly, from −49.8% to −43.1%. However, this performance gain was not sustained at longer lead times. At 36 h and 48 h lead times, NSE dropped to 0.71 and RMSE increased to 193 m³/s, with only modest improvements in bias (Figure 9). Although ensemble predictions outperformed single predictions in both events, the optimal configuration varied. For Event 2020, longer lead time ensembles (Ens-36h and Ens-48h) provided better accuracy and bias correction, particularly at Kangchang. For Event 2022, which featured rapidly evolving meteorological conditions, the mid-range configuration (Ens-24h) proved most effective, while longer configurations showed diminishing performance with increased prediction horizon.

3.2.3. Statistical Analysis of Adaptive Performance

To provide a comprehensive statistical evaluation of forecast performance beyond average metrics, we conducted a detailed analysis specifically at Kangchang station, selected because it represents the integrated downstream hydrological response of the entire Geumho River catchment. Figure 10 presents a quantitative comparison using box plots of three different approaches: single forecasts, ensemble means (Ens-48h), and optimal ensemble configurations that vary by event and lead time. This approach facilitates a comprehensive evaluation of the statistical distribution characteristics of forecast performance beyond the averaged metrics presented in the previous analysis.

The box plot analysis in Figure 10 examines the statistical distribution characteristics of forecast performance across multiple performance criteria. The optimal ensemble configurations shown in Figure 9 were determined through the comprehensive performance evaluation presented in Section 3.2.2, where each lead time window was evaluated against all ensemble configurations capable of providing forecasts for that duration. At 48 h lead time, only Ens-48h can provide forecasts of this duration, precluding optimization comparison.

The optimal ensemble configurations differed depending on the event type; for Event 2020, the Ens-36h configuration was optimal at 12 h and 24 h lead times, whereas Event 2022 performed best with Ens-24h at the same lead times. At the 36 h lead time, both events showed optimal performance with Ens-48h compared to Ens-36h.

The statistical distributions reveal distinct patterns between the two hydrological events across all performance metrics. For the prolonged rainfall event (Event 2020), ensemble forecasts showed narrower interquartile ranges in NSE, KGE, and RMSE, and more centered PBIAS values around zero compared to single forecasts, indicating reduced variability and bias.

The typhoon event (Event 2022) distributions demonstrate higher overall variability compared to Event 2020 across all metrics, confirming the challenging nature of typhoon forecasting. However, the advantages of ensemble approaches become more apparent under these conditions. The optimal configurations show improvements in median performance across NSE, KGE, and RMSE distributions, while PBIAS distributions indicate reduced systematic underestimation bias compared to single forecasts.

Table 3. Median values of forecast performance metrics at Kangchang station by lead time for Events 2020 and 2022 (from Figure 10).

Event	Ens. Configuration	Type	NSE	KGE	RMSE (m3/s)	PBIAS (%)
2020	12 h	Single	0.86	0.89	228.2	−2.1
		Ens-48h	0.88	0.91	216.6	2.1
		Ens-Optimal	0.90	0.92	204.0	0.0
	24 h	Single	0.76	0.83	309.7	−4.2
		Ens-48h	0.87	0.89	229.9	3.0
		Ens-Optimal	0.88	0.90	221.9	1.1
	36 h	Single	0.53	0.70	428.8	−2.3
		Ens-48h	0.83	0.86	256.0	4.4
		Ens-Optimal	0.83	0.86	256.0	4.4
	48 h	Single	0.39	0.57	488.8	0.1
		Ens-48h	0.81	0.83	257.9	5.2
2022	12 h	Single	0.34	0.55	305.6	−57.2
		Ens-48h	0.80	0.69	162.4	−41.0
		Ens-Optimal	0.86	0.70	130.6	−47.1
	24 h	Single	0.48	0.58	271.3	−49.8
		Ens-48h	0.76	0.68	181.0	−39.2
		Ens-Optimal	0.85	0.69	141.3	−43.1
	36 h	Single	0.47	0.57	271.3	−48.9
		Ens-48h	0.71	0.68	188.8	−37.5
		Ens-Optimal	0.71	0.68	188.8	−37.5
	48 h	Single	0.48	0.58	265.3	−45.1
		Ens-48h	0.71	0.68	192.9	−33.7

quantifies the distribution characteristics observed in Figure 10 through median values, revealing the magnitude of performance improvements achieved by optimal ensemble configurations. The statistical analysis demonstrates that ensemble advantages extend beyond simple accuracy improvements to encompass enhanced forecast reliability and reduced prediction uncertainty.

For Event 2020, the combination of Figure 10 and Table 3 reveals consistent performance patterns across lead times. The narrow distribution spreads shown in Figure 10 correspond to stable median performance in Table 3, with optimal configurations achieving NSE values of 0.90 at 12 h and 0.88 at 24 h lead times compared to single forecast values of 0.86 and 0.76, respectively. The PBIAS distributions center near zero for optimal configurations, as confirmed by Table 3 median values of 0.00% and 1.14% compared to negative bias in single forecasts. This pattern indicates that distributed precipitation events benefit from ensemble averaging through both improved accuracy and enhanced consistency.

For Event 2022, Figure 10 and Table 3 together demonstrate the critical value of ensemble approaches under challenging forecast conditions. While typhoon events exhibit higher overall variability, optimal configurations provide substantial improvements in both central tendency and distribution characteristics. The 24 h lead time shows the most pronounced benefits, with NSE improving from 0.48 to 0.85 (Table 3) while Figure 10 reveals dramatic reduction in distribution spread and fewer low-performing outliers. RMSE improvements from 271.32 to 141.26 m³/s (Table 3) correspond to narrower, more reliable distributions in Figure 10.

Cross-metric analysis demonstrated consistent improvements across NSE, RMSE, KGE, and PBIAS, indicating that the ensemble method comprehensively enhances both accuracy and reliability. Figure 10 clearly illustrates improvements achieved by optimal ensemble configurations, reflected in narrower distributions and better central tendency.

The comparison between fixed ensemble approaches (Ens-48h) and optimal configurations highlights the operational value of adaptive strategies. While Ens-48h provides consistent baseline improvements over single forecasts, optimal configurations offer additional gains that are particularly valuable during extreme events. For Event 2022 at 24 h lead time, the difference between Ens-48h (NSE = 0.76) and optimal configuration (NSE = 0.85) represents substantial practical improvement for flood warning applications.

These findings demonstrate that ensemble effectiveness depends not only on averaging multiple forecasts but on selecting appropriate configurations based on meteorological event characteristics.

4. Discussion

Ensemble forecasting methods, including time-lagged techniques, have been extensively validated in operational hydrological systems [30,32]. In particular, time-lagged ensemble forecasting has been successfully applied at continental scales, such as in NOAA’s National Water Model [18], and supported by foundational studies at regional scales [20,21]. Additionally, European flood management systems like EFAS have consistently endorsed the operational advantages of ensemble forecasting [11,24]. Building on this operational context, ensemble streamflow forecasting studies have adopted diverse validation strategies depending on their objectives. Archive-based validation using multi-year hindcasts is commonly used to establish statistical confidence and probabilistic skill metrics for operational reliability assessment [33,34]. In contrast, event-focused validation examines how forecast configuration interacts with storm evolution and meteorological characteristics [20,35]. In this study, we adopt the event-focused approach and systematically evaluate multiple time-lagged configurations (12 h, 24 h, 36 h, 48 h) for two contrasting events, including a prolonged rainfall event (Event 2020) and rapidly evolving typhoon event (Event 2022). This design enables process level interpretation of why different storm evolution patterns favor different ensemble lead times and provides a foundation for future validation using larger event archives.

This study enhances existing knowledge by demonstrating that adapting specific ensemble configurations with distributed hydrological modeling effectively mitigates accuracy degradation observed in single deterministic forecasts, particularly under challenging numerical weather prediction (NWP) conditions. The findings revealed significant improvements in prediction accuracy during critical periods, with relative improvements in Nash–Sutcliffe Efficiency (NSE) reaching up to 108% for prolonged events and 77% for rapid events compared to deterministic forecasts. These performance differences indicate that precipitation forecast uncertainty into streamflow predictions. Figure 6 suggests that time-lagged ensemble averaging moderates runoff prediction uncertainty in an event dependent manner. For Event 2020, longer time-lagged configurations with ensemble lead times of 36 to 48 h stabilize runoff predictions by reducing short term precipitation fluctuations while maintaining multi day accumulation, whereas for Event 2022, the 24 h configuration better retains peak rainfall intensity and the resulting sharp streamflow peaks.

A key motivation for employing the full forecast horizon in time-lagged ensembles is to ensure hydrological predictions extend up to the maximum lead times provided by numerical weather forecasts. However, this study found that adapting ensemble configurations to specific future forecast horizons can achieve notable performance gains without shortening total forecast lead times. Although event-scale precipitation intensity, total precipitation hours, and ensemble spread provide qualitative signals for window selection, deriving an operational decision rule requires validation across a larger and more diverse event set.

This research specifically targeted basin-scale short-range predictions, an area less extensively explored in prior ensemble forecasting studies. Despite limitations arising from the small number of hydrological events analyzed, the results provide essential insights for developing adaptive strategies tailored to specific meteorological conditions. Expanding the analysis to include a broader range of hydrological scenarios would further enhance understanding and enable comprehensive identification of optimal ensemble configurations, representing a promising direction for future research.

The operational implications of these event-dependent and lead-time-dependent variations underscore the importance of dynamic ensemble strategies. Rather than fixed configurations, adaptive methods tailored to anticipated meteorological conditions ensure enhanced accuracy and reliability, particularly during extreme hydrological events. However, operational implementation requires a priori event type classification based on short-range NWP forecasts and substantial computational resources for real-time high-resolution ensemble modeling, challenges that are not fully addressed in this event-based evaluation.

Unlike studies commonly using bias-corrected meteorological inputs [36,37], this research leveraged raw, uncalibrated LDAPS forecasts to directly assess ensemble averaging benefits with distributed hydrological modeling. Despite systematic precipitation underestimation by LDAPS, ensemble averaging effectively compensated for input biases under favorable conditions, emphasizing the potential for further improvements through targeted input calibration or bias correction.

The high-resolution (25 to 100 m scale) implementation effectively captured localized hydrological dynamics, offering practical benefits for regional flood risk management and response planning. Applying the proposed framework to other regions depend on the availability of high-resolution spatial datasets and catchment characteristics, and optimal configurations may require local validation for different basins or NWP models. Parameter uncertainty, while not explicitly quantified here, is another factor that can influence simulated streamflow. Recent studies have shown that optimized parameters can vary with event characteristics and antecedent conditions [38], suggesting that parameter ensemble approaches or alternative calibrated parameter sets warrant consideration in comprehensive operational forecasting systems.

Ultimately, this study highlights that adaptive, basin-scale time-lagged ensemble forecasting strategies can significantly enhance short-range prediction reliability, particularly under challenging meteorological conditions.

5. Conclusions

In this study, we proposed a time-lagged ensemble approach using LDAPS forecasts and the WRF-Hydro model to reduce uncertainty in single short-term streamflow predictions. The proposed method focused on evaluating the predictive capability of different time-lagged ensemble configurations across rainfall events with distinct characteristics. Evaluations were performed on two specific rainfall events (Event 2020 and Event 2022) that occurred in Korea’s Geumho River basin. The impact assessment was conducted using various forecasting methods, including single streamflow forecasts and LDAPS rainfall forecasts categorized by lead time. Additionally, time-lagged ensemble streamflow forecasts with different lead times (12 h, 24 h, 36 h, and 48 h) and model performance metrics were used for evaluation. Based on the results, the following conclusions were drawn:

• Time-lagged ensemble predictions outperformed single predictions across lead times. For Event 2020, the 48 h ensemble improved NSE from 0.39 to 0.81 (108% increase). For Event 2022, the 24 h ensemble raised NSE from 0.48 to 0.85 (77% increase). RMSE decreased by 48% from 271.32 to 141.26 m³/s for Event 2022 at optimal configurations.
• Adaptive selection of ensemble configurations further enhanced prediction accuracy beyond the general improvements achieved by ensemble averaging alone. While fixed configurations like Ens-48h already enhanced performance over single forecasts, optimally selected configurations (e.g., Ens-24h for Event 2022) improved NSE by an additional 11.8%. PBIAS also showed meaningful bias correction, from 49.8% to −43.1% for Event 2022 and approaching zero (0.00% to 1.14%) for Event 2020 under optimal conditions.
• Optimal ensemble configurations depended significantly on event characteristics; Event 2020 performed best under longer lead times (Ens-36h and Ens-48h), whereas Event 2022, influenced by rapidly changing meteorological conditions, showed optimal results at mid-range lead times (Ens-24h). These results highlight the importance of aligning ensemble strategies with the temporal dynamics of rainfall events.

This study highlights the effectiveness of employing a tailored, time-lagged ensemble approach in streamflow forecasting to significantly mitigate predictive uncertainties. The findings underline the necessity of adapting ensemble strategies to specific meteorological contexts to improve forecasting accuracy and reliability, providing a practical pathway for enhancing flood risk management practices.