Multi-step-ahead monthly streamflow forecasting by a neurofuzzy network model

Abstract. Skillful seasonal streamflow forecasts obtained from climate and land surface conditions could significantly improve water and energy management. Since climate forecasts are updated on a monthly basis, we evaluate the potential in developing operational monthly streamflow forecasts on a continuous basis throughout the year. Further, basins in the rainfall–runoff regime critically depend on the forecasted precipitation in the upcoming months as opposed to snowmelt regimes where initial hydrological conditions (IHC) play a critical role. The goal of this study is to quantify the role of updated monthly precipitation forecasts and IHC in forecasting 6-month lead monthly streamflow and soil moisture for a rainfall–runoff mechanism dominated basin – Apalachicola River at Chattahoochee, FL. The Variable Infiltration Capacity (VIC) land surface model is implemented with two forcings: (a) updated monthly precipitation forecasts from ECHAM4.5 Atmospheric General Circulation Model (AGCM) forced with sea surface temperature forecasts and (b) daily climatological ensembles. The difference in skill between the above two quantifies the improvements that could be attainable using the AGCM forecasts. Monthly retrospective streamflow forecasts are developed from 1981 to 2010 and streamflow forecasts estimated from the VIC model are also compared with those predicted by using the principal component regression (PCR) model. The mean square error (MSE) in predicting monthly streamflows, using the VIC model, are compared with the MSE of streamflow climatology under ENSO (El Niño Southern Oscilation) conditions as well as under normal years. Results indicate that VIC forecasts obtained using ECHAM4.5 are significantly better than VIC forecasts obtained using climatological ensembles and PCR models over 2–6 month lead time during winter and spring seasons in capturing streamflow variability and reduced mean square errors. However, at 1-month lead time, streamflow utilizing the climatological forcing scheme outperformed ECHAM4.5 based streamflow forecasts during winter and spring, indicating a dominant role of IHCs up to a 1-month lead time. During ENSO years, streamflow forecasts exhibit better skill even up to a six-month lead time. Comparisons of the seasonal soil moisture forecasts, developed using ECHAM4.5 forcings, with seasonal streamflows also show significant skill, up to a 6-month lead time, in the four seasons.

This study investigates the potential predictability of streamflow and soil moisture in the Alabama–Coosa–Tallapoosa (ACT) river basin in the southeastern United States. The study employs the state-of-the-art National Water Model (NWM) and compares the effects of initial soil moisture condition with those of seasonal climate anomalies on streamflow and soil moisture forecast skills. We have designed and implemented seasonal streamflow forecast ensemble experiments following the methodology suggested by Dirmeyer et al. The study also compares the soil moisture variability in the NWM with in situ measurements and remote sensing data from the Soil Moisture Active and Passive (SMAP) satellite. The NWM skillfully simulates the observed streamflow in the ACT basin. The soil moisture variability is 46% smaller in the NWM compared with the SMAP data, mainly due to a weaker amplitude of the seasonal cycle. This study finds that initial soil moisture condition is a major source of predictability for the seasonal streamflow forecast. The contribution of the initial soil moisture condition is comparable or even higher than that of seasonal climate anomaly effects in dry seasons. In the boreal summer season, the initial soil moisture condition contributes to 65% and 48% improvements in the seasonal streamflow and soil moisture forecast skills, respectively. This study attributes a greater improvement in the streamflow forecast skill to the lag effects between the soil moisture and streamflow anomalies. The results of this study can inform the development and improvement of the operational streamflow forecasting system.

Society’s increasing demand for water and the need for its long-term management have motivated efforts toward improving seasonal streamflow forecasts. Currently, seasonal climate forecasts are routinely issued in meteorological centers around the world, generating information for decision-making and seasonal streamflow forecasting (SSF) studies that are becoming more frequent. Seasonal streamflow forecast skill derives from land surface initial conditions and atmospheric boundary conditions that mostly depend on large-scale climate phenomena (such as ENSO). Thus, seasonal rainfall predictions produced by dynamic climate models that represent ocean-atmosphere interactions may have a positive impact on streamflow forecasts. In South America, seasonal streamflow forecasts are essential for the hydropower sector, which is responsible for ~65% of the electric energy produced in countries such as Brazil. In this work, we assessed seasonal streamflow forecasts over South America based on a continental-scale application of a hydrologic-hydrodynamic model and precipitation forecasts from the ECMWF's fifth generation seasonal forecast system (SEAS5). Seasonal streamflow forecasts (SEAS5-SF) were evaluated against a reference model run and forecast skill was estimated relative to the Ensemble Streamflow Prediction (ESP) method. The bias correction of SEAS5 predicted precipitation improved the performance of the seasonal streamflow forecasts, frequently turning negative skill results into near null to positive skill. Results indicate that the ESP remains a hard-to-beat method for seasonal streamflow forecasting in South America. SEAS5-SF skill was found to be dependent on initialization month, season, basin and forecast lead time, with greater skill on the initialization month lead time. Rivers where the forecast skill is higher were Amazon, Araguaia, Tocantins and Paraná. Acknowledgments: This work presents part of the results obtained during the project granted by the Brazilian Agency of Electrical Energy (ANEEL) under its Research and Development program Project PD 6491-0503/2018 – “Previsão Hidroclimática com Abrangência no Sistema Interligado Nacional de Energia Elétrica” developed by the Paraná State electric company (COPEL GeT), the Meteorological System of Paraná (SIMEPAR) and the RHAMA Consulting company. The Hydraulic Research Institute (IPH) from the Federal University of Rio Grande do Sul (UFRGS) contribute to part of the project through an agreement with the RHAMA company (IAP-001313).

Monthly streamflow forecasting is vital for managing water resources. Recently, numerous studies have explored and evidenced the potential of artificial intelligence (AI) models in hydrological forecasting. In this study, the feasibility of the convolutional neural network (CNN), a deep learning method, is explored for monthly streamflow forecasting. CNN can automatically extract critical features from numerous inputs with its convolution–pooling mechanism, which is a distinct advantage compared with other AI models. Hydrological and large-scale atmospheric circulation variables, including rainfall, streamflow, and atmospheric circulation factors are used to establish models and forecast streamflow for Huanren Reservoir and Xiangjiaba Hydropower Station, China. The artificial neural network (ANN) and extreme learning machine (ELM) with inputs identified based on cross-correlation and mutual information analyses are established for comparative analyses. The performances of these models are assessed with several statistical metrics and graphical evaluation methods. The results show that CNN outperforms ANN and ELM in all statistical measures. Moreover, CNN shows better stability in forecasting accuracy.

[1] Seasonal streamflow forecasts are valuable for planning and allocation of water resources. In Australia, the Bureau of Meteorology employs a statistical method to forecast seasonal streamflows. The method uses predictors that are related to catchment wetness at the start of a forecast period and to climate during the forecast period. For the latter, a predictor is selected among a number of lagged climate indices as candidates to give the “best” model in terms of model performance in cross validation. This study investigates two strategies for further improvement in seasonal streamflow forecasts. The first is to combine, through Bayesian model averaging, multiple candidate models with different lagged climate indices as predictors, to take advantage of different predictive strengths of the multiple models. The second strategy is to introduce additional candidate models, using rainfall and sea surface temperature predictions from a global climate model as predictors. This is to take advantage of the direct simulations of various dynamic processes. The results show that combining forecasts from multiple statistical models generally yields more skillful forecasts than using only the best model and appears to moderate the worst forecast errors. The use of rainfall predictions from the dynamical climate model marginally improves the streamflow forecasts when viewed over all the study catchments and seasons, but the use of sea surface temperature predictions provide little additional benefit.

<p>Seasonal streamflow forecasts represent critical operational inputs for water sectors and society, for instance for spring flood early warning, water supply, hydropower generation, and irrigation scheduling. Initial hydrological conditions (e.g., snow cover and soil moisture) are an important driver of hydrological predictions on these timescales. In high-latitude and/or high-altitude basins across North America, and the basins downstream of these headwaters, snow is one of the main sources of runoff generation. As a result, data-driven forecasting from snow observations is a well-established approach for operational seasonal streamflow forecasting in the USA (Fleming et al., 2021) and Canada (Zahmatkesh et al., 2019).</p><p>As part of the Global Water Futures programme (GWF), we are advancing capabilities for probabilistic streamflow forecasting over North America. The first aim of this work is to benchmark probabilistic seasonal streamflow predictability across the continent. To this end, a data-driven probabilistic seasonal streamflow hindcasting system is being developed and implemented for basins with a nival regime across North America. It uses snow water equivalent measurements from the recent update of the Canadian historical Snow Water Equivalent dataset (CanSWE, 1928–2020; Vionnet et al., 2021) and the Natural Resources Conservation Service (NRCS) manual snow surveys and the SNOTEL automatic snow pillow in the USA. These datasets are gap filled using quantile mapping based on neighbouring snow and precipitation stations (SCDNA dataset; Tang et al., 2020), and subsequently transformed into principal components. These principal components are then used as predictors into a regression model, to generate ensemble hindcasts of streamflow volumes for basins across North America. Preliminary results indicate that this approach is skilful (i.e., better than streamflow climatology) for basins across the Canadian Rockies during the snowmelt season.</p><p><strong>References</strong></p><p>Fleming, S. W., Garen, D. C., Goodbody, A. G., McCarthy, C. S., and Landers, L. C.: Assessing the new Natural Resources Conservation Service water supply forecast model for the American West: A challenging test of explainable, automated, ensemble artificial intelligence. Journal of Hydrology, 602, https://doi.org/10.1016/j.jhydrol.2021.126782, 2021.</p><p>Tang, G., Clark, M. P., Newman, A. J., Wood, A. W., Papalexiou, S. M., Vionnet, V., and Whitfield, P. H.: SCDNA: a serially complete precipitation and temperature dataset for North America from 1979 to 2018, Earth Syst. Sci. Data, 12, 2381–2409, https://doi.org/10.5194/essd-12-2381-2020, 2020.</p><p>Vionnet, V., Mortimer, C., Brady, M., Arnal, L., and Brown, R.: Canadian historical Snow Water Equivalent dataset (CanSWE, 1928–2020), Earth Syst. Sci. Data, 13, 4603–4619, https://doi.org/10.5194/essd-13-4603-2021, 2021.</p><p>Zahmatkesh, Z., Sanjeev Kumar, J., Coulibaly, P., and Stadnyk, T.: An overview of river flood forecasting procedures in Canadian watersheds, Canadian Water Resources Journal / Revue canadienne des ressources hydriques, 44, 3, https://doi.org/10.1080/07011784.2019.1601598, 2019.</p>

Are hybrid models integrated with data preprocessing techniques suitable for monthly streamflow forecasting? Some experiment evidences

<p>Streamflow information for the months ahead is of great value to existing decision-making practices, particularly to those affected by the vagaries of the climate and who would benefit from better understanding and managing climate-related risks. Despite the large effort, there is still limited knowledge of the key drivers controlling the quality of the seasonal streamflow forecasts. In this investigation, we show that the seasonal streamflow predictability can be clustered, and hence regionalised, based on a priori knowledge of local hydro-climatic conditions. To reach these conclusions we analyse the seasonal forecasts of streamflow volumes across about 35400 basins in Europe, which vary in terms of climatology, scale and hydrological regime. We then link the forecast quality to various descriptors including physiography, hydro-climatic characteristics and meteorological biases. This allows the identification of the key drivers along a strong hydro-climatic gradient. Results show that, as expected, the seasonal streamflow predictability varies geographically and seasonally with acceptable values for the first lead months. In addition, the predictability deteriorates with increasing lead months particularly in the winter months. Nevertheless, we show that the forecast quality is well correlated to a set of drivers, which vary depending on the initialization month. The forecast quality of seasonal streamflow volumes is strongly dependent on the basin’s hydrological regime, with quickly reacting basins (of low river memory) showing limited predictability. On the contrary, snow and/or baseflow dominated regions with long recessions (and hence high river memory) show high streamflow predictability. Finally, climatology and precipitation biases are also strongly related to streamflow predictability, highlighting the importance of developing robust bias-adjustment methods.</p>

Use of an Ensemble Re-ordering Method for disaggregation of seasonal categorical rainfall forecasts into conditioned ensembles of daily rainfall for hydrological forecasting

Monthly streamflow forecasting plays an important role in water resources management, especially for dam operation. In this paper, an approach of model fusion technique named selected model fusion (SMF) is applied and assessed under two strategies of model selection in order to improve the accuracy of streamflow forecasting. The two strategies of SMF are: fusion of the outputs of best individual forecasting models (IFMs) selected by dendrogram analysis (S1), and fusion of the best outputs of all IFMs resulting from an ordered selection algorithm (S2). In both strategies, five data-driven models including: artificial neural network, generalized regression neural network, least square-support vector regression, K-nearest neighbor regression, and multiple linear regression with optimized structure are performed as IFMs. The SMF strategies are applied for forecasting the monthly inflow to Karkheh reservoir, Iran, owning various patterns between predictor and predicted variables in different months. Results show that applying SMF approach based on both strategies results in more accurate forecasts in comparison with fusion of all IFMs outputs (S3), as the benchmark. However, comparison of the two SMF strategies reveals that the implementation of strategy (S2) considerably improves the accuracy of forecasts than strategy (S1) as well as the best IFM results (S4) in all months.

Despite recent advances in operational streamflow forecasting systems, anticipating and forecasting droughts and associated low-flow conditions remain major challenges in hydrology, with substantial impacts on water-dependent sectors such as agriculture and industry. Enhancing sub-seasonal to seasonal streamflow forecasting is therefore critical for improving water resources management. This study investigates the performance of a Handoff forecast Long Short-Term Memory (LSTM) architecture (Nearing et al., 2024) for probabilistic streamflow forecasting at lead times extending up to six months, with a particular emphasis on low-flow conditions.The Handoff forecast LSTM is trained regionally on a subset of 292 basins from the Catchment Attributes and MEteorology for Large-sample Studies - FRance dataset (CAMELS-FR) (Delaigue et al., 2025), after excluding basins affected by unreliable low-flow measurements. Model training relies on basin-averaged hydro-meteorological reanalysis data provided by CAMELS-FR. Evaluation of the model is conducted using ensemble streamflow forecasts generated from historical scenarios and meteorological ensemble predictions from the SEAS5 model from the European Center for Medium-Range Weather Forecasts (ECMWF) (Johnson et al., 2019)The generated ensemble streamflow forecasts are evaluated using a set of probabilistic metrics such as the Continuous Ranked Probability Score (CRPS), the Brier Score, the Area under the ROC curve, and the Talagrand diagram, and using the natural streamflow climatology as a reference. In addition, a sensitivity analysis of static catchment attributes is performed to assess their relative contribution to model performance and to better understand the drivers of predictability across basins.Delaigue, O., Guimarães, G. M., Brigode, P., Génot, B., Perrin, C., Soubeyroux, J.-M., Janet, B., Addor, N., & Andréassian, V. (2025). CAMELS-FR dataset: a large-sample hydroclimatic dataset for France to explore hydrological diversity and support model benchmarking. Earth System Science Data, 17(4), 1461–1479. https://doi.org/10.5194/essd-17-1461-2025Johnson, S. J., Stockdale, T. N., Ferranti, L., Balmaseda, M. A., Molteni, F., Magnusson, L., Tietsche, S., Decremer, D., Weisheimer, A., Balsamo, G., Keeley, S. P. E., Mogensen, K., Zuo, H., & Monge-Sanz, B. M. (2019). SEAS5: The new ECMWF seasonal forecast system. Geoscientific Model Development, 12(3), 1087–1117. https://doi.org/10.5194/gmd-12-1087-2019Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T. Y., Weitzner, D., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559–563. https://doi.org/10.1038/s41586-024-07145-

A hybrid approach to monthly streamflow forecasting: Integrating hydrological model outputs into a Bayesian artificial neural network

Streamflow forecasting is crucial for planning, designing, and managing water resources. Accurate streamflow forecasting is essential in developing water resource systems that are both technically and economically efficient. This study tested several machine learning techniques to estimate monthly streamflow data in the Hunza River Basin, Pakistan, using streamflow, precipitation, and air temperature data between 1985 and 2013. The techniques tested included adaptive boosting (AB), gradient boosting (GB), random forest (RF), and K-nearest neighbors (KNN). The models were developed using river discharge as the target variable, while air temperature and precipitation as the input variables. The model's performance was assessed via four statistical performance indicators namely root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). The results obtained for RMSE, MSE, MAE, and R2 using AB, GB, RF, and KNN techniques are (16.8, 281, 6.53, and 0.998), (95.1, 9,047, 61.5, and 0.921), (126.8, 16,078, 74.6, and 0.859), and (219.9, 48,356, 146.3, and 0.775), respectively. The results indicate that AB outperforms GB, RF, and KNN in predicting monthly streamflow for the Hunza River Basin. Machine learning, particularly AB, offers a reliable approach for streamflow forecasting, aiding hazard and water management in the area.

Accurately forecasting streamflow values is essential to achieve an efficient, integrated water resources management strategy and to provide consistent support to water decision-makers. We present a simple, low-cost, and robust approach for forecasting monthly and yearly streamflows during the current hydrological year, which is applicable to headwater catchments. The procedure innovatively combines the use of well-known regression analysis techniques, the two-parameter Gamma continuous cumulative probability distribution function and the Monte Carlo method. Several model performance statistics metrics (including the Coefficient of Determination R 2 ; the Root-Mean-Square Error RMSE; the Mean Absolute Error MAE; the Index of Agreement IOA; the Mean Absolute Percent Error MAPE; the Coefficient of Nash-Sutcliffe Efficiency NSE; and the Inclusion Coefficient IC) were used and the results showed good levels of accuracy (improving as the number of observed months increases). The model forecast outputs are the mean monthly and yearly streamflows along with the 10th and 90th percentiles. The methodology has been successfully applied to two headwater reservoirs within the Guadalquivir River Basin in southern Spain, Water: Ecology and Management www.videleaf.com achieving an accuracy of 92% and 80% in March 2017. These risk-based predictions are of great value, especially before the intensive irrigation campaign starts in the middle of the hydrological year, when Water Authorities have to ensure that the right decision is made on how to best allocate the available water volume between the different water users and environmental needs.

• Existing un-trended climatology forecasts of seasonal streamflow are prone to bias when trends are present in the streamflow record. • Trended climatology forecasts can remove bias, but are usually worse than un-trended climatology forecasts. • Selective adoption of trended climatology forecasts, based on observed trend significance, can lead to meaningful improvements. Probabilistic climatology forecasts of seasonal streamflow up to 12 months ahead are incredibly useful for water resource planning, informing risk-based management strategies. However, the assumption of a stationary distribution for such climatology forecasts fails to account for observed trends present in many streamflow gauge records, which can result in unnecessary bias and excessive spread in the forecasts. For example, where drying trends have led to reduced flow, water resources may be overestimated based on a higher historical mean flow. While trended climatology forecasts are sometimes far better than un-trended climatology forecasts, most of the time they are worse, in part due to the additional parameter uncertainty from trend estimation. Here we present an approach to selectively adopt trended climatology forecasts based on observed trends in the training period. We use a high-quality hydrological station network covering a broad range of hydroclimate conditions to compare trended and un-trended climatology and evaluate our proposed approach to selective adoption of trended climatology. We found trended climatology forecasts provided high skill relative to the un-trended climatology baseline when negative trends were present, especially when those trends were significant. We also observed negative skill in around half the cases, especially where positive trends were present, highlighting the necessity for selective application. We recommend the adoption of trended climatology for seasonal streamflow forecasting whenever there is a fair chance of a negative trend present (and not when a positive trend is present), particularly in applications that require conservative estimates of long-term flows (e.g., allocation of annual water accounts).