Longer Lead Times Research Articles

RUFCO: a deep-learning framework to post-process subseasonal precipitation accumulation forecasts

Abstract Post-processing is a critical step in attaining calibrated and reliable probabilistic forecast output from numerical weather prediction models. A novel deep-learning framework is proposed to post-process 20 years of 7- and 14-day precipitation accumulation reforecasts from the Global Ensemble Forecast System at subseasonal timescales (week 1, week 2, and combined weeks 3-4 forecasts) over the contiguous United States. The network builds upon previous studies and is a combination of three parallel-trained components suitable for subseasonal prediction. The first is a ResUnet architecture which learns non-linear relationships between binned observed precipitation and input images of weather and geographical variables. The second conditions the network to the month-of-year via a Feature-wise Linear Modulation (FiLM) layer. The third helps the network learn when to revert the forecast to that of climatology. The RUFCO (named for its components ResUnet, FiLM, and Climatological-Offramp) forecasts are compared against raw and climatological forecasts as well as those from a state-of-the-art distributional regression post-processing model, “CSGD,” and a simple bias-corrected model. At week 1, every method exhibited a competitive advantage over climatological forecasts. At week 2, RUFCO generated forecasts with statistically significant improvement over climatology at 82-94% of the domain, beating CSGD’s coverage of 76-90% of the grid points. At week 3, RUFCO’s skillful coverage was 65-85% while CSGD’s dropped to only 12-37%. At the longer lead times, RUFCO achieved the highest domain-averaged skill scores across seasons. However, the network tends to “smooth” forecast skill making it less competitive with CSGD in limited areas with strongly spatially-varying biases.

Atmospheric Precursors of Skillful SST Prediction in the Northeast Pacific

Abstract Forecasts of sea surface temperature anomalies (SSTAs) provide essential information to stakeholders of marine resources in coastal ecosystems, such as the California Current Large Marine Ecosystem (CCLME), at management-relevant monthly-to-annual time scales. Diagnosing dynamical sources of predictability and the mechanisms differentiating skill among forecasts is required for verification and improvement in operational forecasting systems. Using retrospective forecasts (1982–2020) from a four-member subset of the North American Multi-Model Ensemble (NMME), we evaluate the conditional skill of SSTA forecasts in the CCLME at monthly resolution for lead times up to 10.5 months. Forecasts from ensemble members with relatively small SSTA errors at shorter lead times retain higher skill at longer lead times, with the most substantial and long-lasting increases for forecasts initialized in the fall and early spring. The “best” low-error SSTA forecasts are characterized by increased skill in the prediction of North Pacific atmospheric circulation [sea level pressure (SLP) and 200-hPa geopotential height] the month prior to the evaluation of SSTA errors in the CCLME and exhibit more realistic progressions of anomalous SLP. The Pacific meridional mode (PMM) emerges as a diagnostic of skillful North Pacific atmosphere–ocean coupling, as forecasts that correctly simulate the PMM and its associated SLP variability increase the SSTA prediction skill in the CCLME in the fall through spring. Predictable coupled ocean–atmosphere modes provide a target for enhancing predictability with early detection of the onset of a deterministic progression emerging from stochastic atmospheric variability. Significance Statement Global forecast systems provide near-term climate predictions that inform the management of marine resources, such as those of the California Current Large Marine Ecosystem. In this study, we probe the processes which lead forecasts to succeed or fail at predicting sea surface temperatures in the California Current at seasonal time scales among retrospective forecasts from the North American Multimodel Ensemble. We demonstrate that forecasts which best simulate sea surface temperatures at the earliest lead times sustain advantages in forecast skill and find that correctly simulating extratropical atmospheric circulation increases the predictive skill of sea surface temperatures in the northeast Pacific in the following lead times. Our results offer North Pacific atmospheric circulation as a target for forecast model improvement that would additionally enhance ocean forecasts.

Significant winter Atlantic Niño effect on ENSO and its future projection

The Atlantic Niño, a primary climatic variability mode in the equatorial Atlantic Ocean, exhibits pronounced variability not only in boreal summer but also in winter. However, the role of the winter Atlantic Niño in trans-basin interactions remains underexplored compared to its summer counterpart. Through analysis of observational reanalysis data since the mid-twentieth century, here we found that the winter Atlantic Niño significantly influences the development of El Niño–Southern Oscillation (ENSO), surpassing the impact of summer Atlantic Niño, with a longer lead time. This effect is reasonably captured in the CMIP6 Historical simulations from a multi-model ensemble perspective. Further analysis of the global warming scenario projects that the influence of winter Atlantic Niño on ENSO will persist into the future, in contrast to the reduced impact of summer Atlantic Niño. Therefore, these findings underscore the importance of further investigating the winter Atlantic Niño to gain a comprehensive understanding of trans-basin interactions and their future changes.

Development of a Two-Step EOF Statistical Postprocessing Algorithm to Identify Patterns of Systematic Error and Variance Within GEFSv12 Reforecasts

Abstract Analysis of seasonally varying signals of model error biases, as well as the diagnosis of space-time patterns of synoptic error, here is accomplished through the development of a statistical postprocessing algorithm based on time-extended Empirical Orthogonal Functions (EEOFs) built on patterns of variance. This work analyzes GEFSv12 200 hPa Geopotential Height (Z200) model error against ERA5 reanalysis data. The mean squared error variance between GEFSv12 reforecast and ERA5 grows rapidly after day 7 of a forecast, and continues to increase through the end of the 16 day forecast period. At lead time 16, the largest variance occurs in middle to high latitude oceanic storm tracks. Variance is highest during hemispheric winter, when baroclinic energy is abundant. The seasonal cycle of error shows largest anomalies in the high latitudes during hemispheric winter. After running the titular two-step, space-time EEOF algorithm, a standardized spatial eigenspectrum shows eigenvalues increase at longer lead times after decreasing in the medium range, demonstrating that the algorithm extracts large-scale signals of systematic error in the long range. Leading wintertime space-time eigenmode pairs include high latitude blocking structures and Rossby wave trains. Results suggest that the large-scale systematic errors in GEFSv12 Z200 initiate in part from inaccurate representations of phase speeds of atmospheric waves in the polar jet.

Long-range hydrological drought forecasting using multi-year cycles in the North Atlantic Oscillation

With global temperatures, populations and ecological stressors expected to rise, hydrological droughts are projected to have progressively severe economic and environmental impacts. As a result, hydrological drought forecasting systems have become increasingly important water resource management tools for mitigating these impacts. However, high frequency behaviours in meteorological or atmospheric conditions often limit the lead times of hydrological drought forecasts to seasonal timescales, either through poorer performance of multi-year meteorological forecasts or the lack of multi-year lags in atmosphere-hydrology systems. By contrast, low frequency behaviours in regionally important teleconnection systems (such as the North Atlantic Oscillation, NAO) offer a novel way to forecast hydrological drought at longer lead times. This paper shows that, by using a data-driven modelling approach, long-term behaviours within the NAO can be skilful predictors of hydrological drought conditions at a four-year forecasting horizon. Multi-year semi-periodic patterns in the NAO were used to forecast regional groundwater drought coverage in the UK (proportion of groundwater boreholes in drought), with the greatest forecast performance achieved for longer duration droughts, and for hydrogeological regions with longer response times. Model errors vary from 14 % (proportion of boreholes, (MAE)) in flashy hydrological regions or short droughts (<3 months), to 2 % for longer duration droughts (>8 months). Model fits of r2 up to 0.8 were produced between simulated and recorded regional drought coverage. As such our results show that teleconnection indices can be a skilful predictor of hydrological drought dynamics at multi-year timescales, opening new opportunities for long-lead groundwater drought forecasts to be integrated within existing drought management strategies in Europe and beyond.

Listening to Stakeholders III: Potential Users Evaluate Product Content and Design for Subseasonal Extreme Precipitation Forecasts

Abstract Extreme precipitation over a 2-week period can cause significant impacts to life and property. Trustworthy and easy-to-understand forecasts of these extreme periods on the subseasonal-to-seasonal time frame may provide additional time for planning. The Prediction of Rainfall Extremes at Subseasonal to Seasonal Periods (PRES2iP) project team conducted three workshops over 6 years to engage with stakeholders to learn what is needed for decision-making for subseasonal precipitation. In this study, experimental subseasonal-to-seasonal (S2S) forecast products were designed, using knowledge gained from previous stakeholder workshops, and shown to decision-makers to evaluate the products for two 14-day extreme precipitation period scenarios. Our stakeholders preferred a combination of products that covered the spatial extent, regional daily values, with associated uncertainty, and text narratives with anticipated impacts for planning within the S2S time frame. When targeting longer extremes, having information regarding timing of expected impacts was seen as crucial for planning. We found that there is increased uncertainty tolerance with stakeholders when using products at longer lead times that typical skill metrics, such as critical success index or anomaly correlation coefficient, do not capture. Therefore, the use of object-oriented verification that allows for more flexibility in spatial uncertainty might be beneficial for evaluating S2S forecasts. These results help to create a foundation for the design, verification, and implementation of future operational forecast products with longer lead times, while also providing an example for future workshops that engage both researchers and decision-makers.

Subseasonal Forecast Skill of Evaporative Demand, Soil Moisture, and Flash Drought Onset from Two Dynamic Models over the Contiguous United States

Abstract Flash droughts are rapidly developing subseasonal climate extreme events that are manifested as suddenly decreased soil moisture, driven by increased evaporative demand and/or sustained precipitation deficits. Over each climate region in the contiguous United States (CONUS), we evaluated the forecast skill of weekly root-zone soil moisture (RZSM), evaporative demand (ETo), and relevant flash drought (FD) indices derived from two dynamic models [Goddard Earth Observing System model V2p1 (GEOS-V2p1) and Global Ensemble Forecast System version 12 (GEFSv12)] in the Subseasonal Experiment (SubX) project between years 2000 and 2019 against three reference datasets: Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2), North American Land Data Assimilation System, phase 2 (NLDAS-2), and GEFSv12 reanalysis. The ETo and its forcing variables at lead week 1 have moderate-to-high anomaly correlation coefficient (ACC) skill (∼0.70–0.95) except downwelling shortwave radiation, and by weeks 3–4, predictability was low for all forcing variables (ACC &lt; 0.5). RZSM (0–100 cm) for model GEFSv12 showed high skill at lead week 1 (∼0.7–0.85 ACC) in the High Plains, West, Midwest, and South CONUS regions when evaluated against GEFSv12 reanalysis but lower skill against MERRA-2 and NLDAS-2 and ACC skill are still close to 0.5 for lead weeks 3–4, better than ETo forecasts. GEFSv12 analysis has not been evaluated against in situ observations and has substantial RZSM anomaly differences when compared to NLDAS-2, and our analysis identified GEFSv12 reforecast prediction limit, which can maximally achieve ACC ∼0.6 for RZSM forecasts between lead weeks 3 and 4. Analysis of major FD events reveals that GEFSv12 reforecast inconsistently captured the correct location of atmospheric and RZSM anomalies contributing to FD onset, suggesting the needs for improving the dynamic models’ assimilation and initialization procedures to improve subseasonal FD predictability. Significance Statement Flash droughts are rapidly developing climate extremes which reduce soil moisture through enhanced evaporative demand and precipitation deficits, and these events can have large impacts on the ecosystem and crop health. We evaluated the subseasonal forecast skill of soil moisture and evaporative demand against three reanalysis datasets and found that evaporative demand skill was similar between forecasts and reanalyses while soil moisture skill is dependent on the reference dataset. Skill of evaporative demand decreases rapidly after week 1, while soil moisture skill declines more slowly after week 1. Case studies for the 2012, 2017, and 2019 United States flash droughts identified that forecasts could capture rapid decreases in soil moisture in some regions but not consistently, implying that long-lead forecasts still need improvements before being used in early warning systems. Improvements in flash drought predictability at longer lead times will require less biased initial conditions, better model parameterizations, and improved representations of large-scale teleconnections.

Optimized Parameter Estimation and Integrating Neural Network Forecasting of Dynamic Plant-Livestock Model for Early Warning in Agro-Environment Control Systems

The research utilizes the Lotka-Volterra prey-predator model to study Plant-Herbivore dynamics, focusing on the relationship between traditional livestock farming and vegetation conditions. Advanced methods are developed to improve the precision and efficiency of parameter estimation in these models. Neural networks are incorporated to enhance forecasting abilities, and an extension of the Plant-Herbivore models includes Botswana's climate and livestock variables. Efficient parameter space exploration is achieved using the Runge-Kutta method along with Multistart and the local solver $fmincon$ in MATLAB. This method improves parameter estimation accuracy. To address the impact of homogeneity assumptions in the data, estimate aggregation through weighting and time conversion is applied. Furthermore, the study investigates the use of nonlinear least squares to further refine the process, allowing for the identification of parameters that best fit observed livestock data, even with non-linearity. By using optimized parameter estimation techniques along with normalized nonlinear least squares, the cumulative error was reduced from an initial 1563.4521 to a final value of 0.0038, well within the specified thresholds (1.0, 0.1, and 0.01). Comparisons between Autoregressive Integrated Moving Average (ARIMA) and Neural Network Auto-Regressive (NNAR) models showed that NNAR models outperformed ARIMA models, with lower variance estimates (0.000004 - 0.000562) compared to ARIMA (0.103 - 0.155). NNAR models displayed Mean Error (ME) values ranging from -0.0012 to 0.0140, indicating a close match between forecasts and actual values with minor deviations. As a result, NNAR forecasting was used for predicting soil moisture, death, and harvest rates, which were integrated into the extended Plant-Herbivore model. This integration enabled the estimation of livestock production trajectories for 2021-2022, along with corresponding interpretations. The study also assessed the uncertainty propagation from NNAR forecasts onto the Plant-Herbivore dynamic model, revealing an increase in uncertainty with longer lead times.

Strong El Niño Events Lead to Robust Multi‐Year ENSO Predictability

AbstractThe El Niño‐Southern Oscillation (ENSO) phenomenon—the dominant source of climate variability on seasonal to multi‐year timescales—is predictable a few seasons in advance. Forecast skill at longer multi‐year timescales has been found in a few models and forecast systems, but the robustness of this predictability across models has not been firmly established owing to the cost of running dynamical model predictions at longer lead times. In this study, we use a massive collection of multi‐model hindcasts performed using model analogs to show that multi‐year ENSO predictability is robust across models and arises predominantly due to skillful prediction of multi‐year La Nina events following strong El Niño events.

Modelling and analysis of production management system using vacation queueing theoretic approach

This paper explores a queueing model in a production management context, featuring periods of working vacations and Bernoulli vacation. When there are no pending orders, the manufacturing unit transits into a maintenance phase, also termed as working vacation, during which production continues, albeit at a slower rate. This diminished productivity could result in longer lead times and potential customer dissatisfaction. Depending on the influx of demand, the maintenance phase could be interrupted, propelling the unit back to full operational capacity. Conversely, if no orders are received during the maintenance phase, the unit has the choice of switching to standby mode in anticipation of incoming orders, or initiating an extended break. We utilize mathematical techniques such as continued fractions, Modified Bessel functions, and Laplace transforms to precisely compute the transient state probabilities of this model. To further illustrate the impact of these operational dynamics on production management, we present corroborative numerical examples.

Global Precipitation Nowcasting of Integrated Multi-satellitE Retrievals for GPM: A U-Net Convolutional LSTM Architecture

Abstract This paper presents a deep supervised learning architecture for 30-min global precipitation nowcasts with a 4-h lead time. The architecture follows a U-Net structure with convolutional long short-term memory (ConvLSTM) cells empowered by ConvLSTM-based skip connections to reduce information loss due to the pooling operation. The training uses data from the Integrated Multi-satellitE Retrievals for GPM (IMERG) and a few key drivers of precipitation from the Global Forecast System (GFS). The impacts of different training loss functions, including the mean-squared error (regression) and the focal loss (classification), on the quality of precipitation nowcasts are studied. The results indicate that the regression network performs well in capturing light precipitation (&lt;1.6 mm h−1), while the classification network can outperform the regression counterpart for nowcasting of high-intensity precipitation (&gt;8 mm h−1), in terms of the critical success index (CSI). It is uncovered that including the forecast variables can improve precipitation nowcasting, especially at longer lead times in both networks. Taking IMERG as a relative reference, a multiscale analysis, in terms of fractions skill score (FSS), shows that the nowcasting machine remains skillful for precipitation rate above 1 mm h−1 at the resolution of 10 km compared to 50 km for GFS. For precipitation rates greater than 4 mm h−1, only the classification network remains FSS skillful on scales greater than 50 km within a 2-h lead time. Significance Statement This study presents a deep neural network architecture for global precipitation nowcasting with a 4-h lead time, using sequences of past satellite precipitation data and simulations from a numerical weather prediction model. The results show that the nowcasting machine can improve short-term predictions of high-intensity global precipitation. The research outcomes will enable us to expand our understanding of how modern artificial intelligence can improve the predictability of extreme weather and benefit flood early warning systems for saving lives and properties.

Development of Indian summer monsoon precipitation biases in two seasonal forecasting systems and their response to large-scale drivers

Abstract. The Met Office Global Coupled Model (GC) and the National Centers for Environmental Prediction (NCEP) Climate Forecast System (CFSv2) are both widely used for predicting and simulating the Indian summer monsoon (ISM), and previous studies have demonstrated similarities in the biases in both systems at a range of timescales from weather forecasting to climate simulation. In this study, ISM biases are studied in seasonal forecasting setups of the two systems in order to provide insight into how they develop across timescales. Similarities are found in the development of the biases between the two systems, with an initial reduction in precipitation followed by a recovery associated with an increasingly cyclonic wind field to the north-east of India. However, this occurs on longer timescales in CFSv2, with a much stronger recovery followed by a second reduction associated with sea surface temperature (SST) biases so that the bias at longer lead times is of a similar magnitude to that in GC. In GC, the precipitation bias is almost fully developed within a lead time of just 8 d, suggesting that carrying out simulations with short time integrations may be sufficient for obtaining substantial insight into the biases in much longer simulations. The relationship between the precipitation and SST biases in GC seems to be more complex than in CFSv2 and differs between the early part of the monsoon season and the later part of the monsoon season. The relationship of the bias with large-scale drivers is also investigated, using the boreal summer intraseasonal oscillation (BSISO) index as a measure of whether the large-scale dynamics favour increasing, active, decreasing or break monsoon conditions. Both models simulate decreasing conditions the best and increasing conditions the worst, in agreement with previous studies and extending these previous results to include CFSv2 and multiple BSISO cycles.

Value of long-term inflow forecast for hydropower operation: A case study in a low forecast precision region

Many studies have evaluated the value of long-term inflow forecast for hydropower operation in high-precision watersheds and explored the impact of forecast uncertainty on hydropower benefits. However, few of them focuses on the low-accuracy regions and the specific impact of forecast errors on hydropower benefits. Taking Hunjiang cascaded system in China as an example, this study investigated the value of low accuracy long-term inflow forecast for hydropower operation, revealed the influence mechanism of forecast errors on hydropower benefits, and determined the critical threshold for accuracy of beneficial forecast. Results show that low accuracy forecast but higher than critical threshold can increase hydropower generation. Hunjiang has a low accuracy long-term forecast with qualified rate being 30–40 %, but annual hydropower generation can be increased by 14.86–29.58 million kWh. The influence mechanism of forecast errors on hydropower operation presents three typical scenarios, two of which are harmless. One is that the error tolerance of operation rules can avoid some decision-making errors; the other one is that when current misreporting is consistent with the inflow in a longer lead time, the misreporting may instead be beneficial. We conclude that low accuracy long-term forecast is valuable but higher accuracy brings higher benefits.

Research on Frequency Matching Correction Techniques for South China Precipitation Ensemble Forecast Based on the GRAPES Model

This study focuses on the real-time precipitation forecast products of the GRAPES_MESO regional ensemble forecast model, which is developed by the Numerical Weather Prediction Center of the China Meteorological Administration and is initialized 1–3 days in advance at 12:00 UTC. Using a national-level homogenized precipitation grid dataset from surface meteorological stations as observational data, a frequency matching method (FMM) is employed to correct precipitation forecasts for different precipitation intensity levels, including light rain, moderate rain, heavy rain, and torrential rain. Case studies and statistical tests (TS scores) are conducted to compare the forecast performance before and after correction. The results indicate that the model’s Cumulative Distribution Function (CDF) curves deviate from observations, and the longer the lead time, the more significant the error. The correction coefficients (CCs) show an increasing trend with the growth of precipitation intensity, indicating that for larger precipitation amounts and longer lead times, larger CCs are needed, highlighting the necessity of correction. Analyzing two precipitation events in South China in July 2019, the FMM results in an increase in precipitation intensity and a widening of the range of heavy precipitation. The corrected precipitation magnitudes are closer to the observations. The statistical tests using TS scores reveal that the FMM has a certain correction effect on the overall precipitation forecast in the South China region, especially for longer lead times and higher precipitation intensities, where the correction effect is more significant. The necessity of frequency matching correction becomes more apparent for heavier precipitation, and the correction effect becomes more significant with longer lead times.

Skillful prediction of length of day one year ahead in multiple decadal prediction systems

Despite a small amplitude, Length of Day (LOD) change, which varies from one year to another due to changes in Atmospheric Angular Momentum (AAM), determines the accuracy of Global Positioning System (GPS) time calculation. In this study, we examine the prediction skill of LOD and AAM in nine decadal prediction systems archived for the Decadal Climate Prediction Project. A persistence and rebound in LOD prediction skill at one year or longer lead time is found in most models. A poleward propagation of AAM anomaly via wave-mean flow interaction is also qualitatively well reproduced. This long-lead prediction of LOD and AAM is attributed to reliable predictions of the El Niño–Southern Oscillation (ENSO) and the Quasi-Biennial Oscillation (QBO), the former being more systematically related than the latter. This result indicates that the improved ENSO prediction and atmospheric wave-mean flow interaction may lead to better prediction of LOD, AAM and related extratropical climate in the decadal prediction systems.

Comparison of Clustering Approaches in a Multimodel Ensemble for U.S. East Coast Cold Season Extratropical Cyclones

Abstract Several clustering approaches are evaluated for 1–9-day forecasts using a multimodel ensemble that includes the GEFS, ECMWF, and Canadian ensembles. Six clustering algorithms and three clustering spaces are evaluated using mean sea level pressure (MSLP) and 12-h accumulated precipitation (APCP) for cool-season extratropical cyclones across the Northeast United States. Using the MSLP cluster membership to obtain the APCP clusters is also evaluated, along with applying clustering determined at one lead time to cluster forecasts at a different lead time. Five scenarios from each clustering algorithm are evaluated using displacement and intensity/amount errors from the scenario nearest to the MSLP and 12-h APCP analyses in the NCEP GFS and ERA5, respectively. Most clustering strategies yield similar improvements over the full ensemble mean and are similar in probabilistic skill except that 1) intensity displacement space gives lower MSLP displacement and intensity errors; and 2) Euclidean space and agglomerative hierarchical clustering, when using either full or average linkage, struggle to produce reasonably sized clusters. Applying clusters derived from MSLP to 12-h APCP forecasts is not as skillful as clustering by 12-h APCP directly, especially if several members contain little precipitation. Use of the same cluster membership for one lead time to cluster the forecast at another lead time is less skillful than clustering independently at each forecast lead time. Finally, the number of members within each cluster does not necessarily correspond with the best forecast, especially at the longer lead times, when the probability of the smallest cluster being the best scenario was usually underestimated. Significance Statement Numerical weather prediction ensembles are widely used, but more postprocessing tools are necessary to help forecasters interpret and communicate the possible outcomes. This study evaluates various clustering approaches, combining a large number of model forecasts with similar attributes together into a small number of scenarios. The 1–9-day forecasts of both sea level pressure and 12-h precipitation are used to evaluate the clustering approaches for a large number of U.S. East Coast winter cyclones, which is an important forecast problem for this region.

Hydrometeorological Insights into the Forecasting Performance of Multi-Source Weather over a Typical Hill-Karst Basin, Southwest China

Reliable precipitation forecasts are essential for weather-related disaster prevention and water resource management. Multi-source weather (MSWX), a recently released ensemble meteorological dataset, has provided new opportunities with open access, fine horizontal resolution (0.1°), and a lead time of up to seven months. However, few studies have comprehensively evaluated the performance of MSWX in terms of precipitation forecasting and hydrological modeling, particularly in hill-karst basins. The key concerns and challenges are how precipitation prediction performance relates to elevation and how to evaluate the hydrologic performance of MSWX in hill-karst regions with complex geographic heterogeneity. To address these concerns and challenges, this study presents a comprehensive evaluation of MSWX at the Chengbi River Basin (Southwest China) based on multiple statistical metrics, the Soil and Water Assessment Tool (SWAT), and a multi-site calibration strategy. The results show that all ensemble members of MSWX overestimated the number of precipitation events and tended to have lower accuracies at higher altitudes. Meanwhile, the error did not significantly increase with the increased lead time. The “00” member exhibited the best performance among the MSWX members. In addition, the multi-site calibration-enhanced SWAT had reliable performance (Average Nash–Sutcliffe value = 0.73) and hence can be used for hydrological evaluation of MSWX. Furthermore, MSWX achieved satisfactory performance (Nash–Sutcliffe value &gt; 0) in 22% of runoff event predictions, but the error increased with longer lead times. This study gives some new hydrometeorological insights into the performance of MSWX, which can provide feedback on its development and applications.

Advancing very short-term rainfall prediction with blended U-Net and partial differential approaches

Accurate and timely prediction of short-term rainfall is crucial for reducing the damages caused by heavy rainfall events. Therefore, various precipitation nowcasting models have been proposed. However, the performance of these models still remains limited. In particular, the current operational precipitation nowcasting method, which is based on radar echo tracking, such as the McGill Algorithm for Precipitation Nowcasting by Lagrangian Extrapolation (MAPLE), has a critical drawback when predicting newly developed or decayed precipitation fields. Recently proposed deep learning models, such as the U-Net convolutional neural network outperform the models based on radar echo tracking. However, these models are unsuitable for operational precipitation nowcasting due to their blurry predictions over longer lead times. To address these blurry effects and enhance the performance of U-Net-based rainfall prediction, we propose a blended model that combines a partial differential equation (PDE) model based on fluid dynamics with the U-Net model. The evaluation of the forecast skill, based on both qualitative and quantitative methods for 0–3-h lead times, demonstrates that the blended model provides less blurry and more accurate rainfall predictions compared with the U-Net and partial differential equation models. This indicates the potential to enhance the field of very short-term rainfall prediction. Additionally, we also evaluated the monthly-averaged forecast skills for different seasons, and confirmed the operational feasibility of the blended model, which contributes to the performance enhancement of operational nowcasting.

A deep learning modeling framework with uncertainty quantification for inflow-outflow predictions for cascade reservoirs

Accurate prediction of reservoir inflows and outflows and their uncertainties is essential for managing water resources and establishing early-warning systems. However, this can be a formidable challenge due to numerous uncertainties, particularly in cascade reservoir systems. To seamlessly quantify aleatoric (data-caused) and epistemic (model network–caused) uncertainties simultaneously in a single framework, we estimated the posterior distribution of parameters in a Bayesian neural network while measuring prediction variance to reflect the noise in data. By randomly discarding certain units within the network, the estimated posterior distribution can be combined with a new loss function. This sophisticated deep learning framework for a long short-term memory network has also included advanced supporting approaches, such as input variable selection, data transformation, and hyperparameter optimization. The model trained in this study was applied to two cascade reservoir systems with four reservoirs in Vietnam, providing uncertainty estimates from two distinct sources. Comparing three global reservoir operation schemes, we found that the proposed model achieved superior performance in all cases, and uncertainty in the forecasts was due primarily to data noise, rather than uncertainty in the model itself. Preprocessing data using a wavelet transform can reduce noise that the model cannot classify on its own, resulting in improved performance, particular with longer lead times. The satisfactory performance of the prediction results confirms that the framework can effectively assess the uncertainty of hydrologic predictions using deep learning.

Deep learning model based on multi-scale feature fusion for precipitation nowcasting

Abstract. Forecasting heavy precipitation accurately is a challenging task for most deep learning (DL)-based models. To address this, we present a novel DL architecture called “multi-scale feature fusion” (MFF) that can forecast precipitation with a lead time of up to 3 h. The MFF model uses convolution kernels with varying sizes to create multi-scale receptive fields. This helps to capture the movement features of precipitation systems, such as their shape, movement direction, and speed. Additionally, the architecture utilizes the mechanism of discrete probability to reduce uncertainties and forecast errors, enabling it to predict heavy precipitation even at longer lead times. For model training, we use 4 years of radar echo data from 2018 to 2021 and 1 year of data from 2022 for model testing. We compare the MFF model with three existing extrapolative models: time series residual convolution (TSRC), optical flow (OF), and UNet. The results show that MFF achieves superior forecast skills with high probability of detection (POD), low false alarm rate (FAR), small mean absolute error (MAE), and high structural similarity index (SSIM). Notably, MFF can predict high-intensity precipitation fields at 3 h lead time, while the other three models cannot. Furthermore, MFF shows improvement in the smoothing effect of the forecast field, as observed from the results of radially averaged power spectral (RAPS). Our future work will focus on incorporating multi-source meteorological variables, making structural adjustments to the network, and combining them with numerical models to further improve the forecast skills of heavy precipitations at longer lead times.