Low-flow Forecasting Research Articles

Assessing Objective Functions in Streamflow Prediction Model Training Based on the Naïve Method

Reliable streamflow forecasting is a determining factor for water resource planning and flood control. To better understand the strengths and weaknesses of newly proposed methods in streamflow forecasting and facilitate comparisons of different research results, we test a simple, universal, and efficient benchmark method, namely, the naïve method, for short-term streamflow prediction. Using the naïve method, we assess the streamflow forecasting performance of the long short-term memory models trained with different objective functions, including mean squared error (MSE), root mean squared error (RMSE), Nash–Sutcliffe efficiency (NSE), Kling–Gupta efficiency (KGE), and mean absolute error (MAE). The experiments over 273 watersheds show that the naïve method attains good forecasting performance (NSE > 0.5) in 88%, 65%, and 52% of watersheds at lead times of 1 day, 2 days, and 3 days, respectively. Through benchmarking by the naïve method, we find that the LSTM models trained with squared-error-based objective functions, i.e., MSE, RMSE, NSE, and KGE, perform poorly in low flow forecasting. This is because they are more influenced by training samples with high flows than by those with low flows during the model training process. For comprehensive short-term streamflow modeling without special demand orientation, we recommend the application of MAE instead of a squared-error-based metric as the objective function. In addition, it is also feasible to perform logarithmic transformation on the streamflow data. This work underscores the critical importance of appropriately selecting the objective functions for model training/calibration, shedding light on how to effectively evaluate the performance of streamflow forecast models.

Operational low-flow forecasting using LSTMs

This study focuses on exploring the potential of using Long Short-Term Memory networks (LSTMs) for low-flow forecasting for the Rhine River at Lobith on a daily scale with lead times up to 46 days ahead. A novel LSTM-based model architecture is designed to leverage both historical observation and forecasted meteorological data to carry out multi-step discharge time series forecasting. The feature and target selection for this deep learning (DL) model involves evaluating the use of different spatial resolutions for meteorological forcing (basin-averaged or subbasin-averaged), the impact of incorporating past discharge observations, and the use of different target variables (discharge Q or time-differenced discharge dQ). Then, the model is trained using the ERA5 dataset as meteorological forcing, and employed for operational forecast with ECMWF seasonal forecast (SEAS5) data. The forecast results are compared to a benchmark process-based model, wflow_sbm. This study also explores the flexibility of the DL model by fine-tuning the pretrained model with limited SEAS5 dataset. Key findings from feature and target selection include: (1) opting for subbasin-averaged meteorological variables significantly improves model performance compared to a basin-averaged approach. (2) Utilizing dQ as the target variable greatly boosts short-term forecast accuracy compared to using Q, with a mean absolute error (MAE) of 25 m3 s−1 and mean absolute percentage error (MAPE) of 0.02 for the first lead time, ensuring reliability and accuracy at the onset of the forecast horizon. (3) While incorporating historical discharge improves the forecasting of Q, its impact on predicting dQ is less pronounced for short lead times. In the operational forecast with SEAS5, compared to the wflow_sbm model, the DL model exhibits skill in forecasting low flows as evidenced by Continuous Ranked Probability Skill Score (CRPSS) median values of all lead times above zero, and better accuracy in forecasting drought events within short lead times. The wflow_sbm model shows higher accuracy for longer lead times. In the exploration of fine-tuning approach, the fine-tuned model generates marginal short-term enhancements in forecasting low-flow events over a non-fine-tuned model. Overall, this study contributes to advancing the field of low-flow forecasting using deep learning approach.

Performance Analysis of Baseflow Separation Methods: The Case of Rift Valley Lakes Basin, Ethiopia

Adopting the appropriate method to separate baseflow from stream flow is desirable for future low flow prediction, planning, management of water resources, and nourishing the environment as well. Thus, comparing the baseflow separation method is inevitable unfortunately not studied within the basin. Therefore, in this study, seven recursive digital filters (RDF) and two digital graphical (DGM) methods were compared in rift valley lakes basins. All the methods were calibrated manually with the help of BFI 3.0 tool; the performance of each method was checked by R2 and RMSE, taking the separation with maximum R2 and minimum RMSE were taken as appropriate separation method and (Baseflow Index) BFI was calculated by using the baseflow from the suitable method for each catchment. The outcomes of baseflow separation indicate that two methods (exponentially weighted moving average (EWMA) and Lynie-Holick) performed better than the other seven methods; unlikely, local minimum and one parameter methods perform less by both R2 and RMSE. Therefore, these comparisons could possibly elucidate the baseflow prediction in the majority of catchments. Subsequently, existing and forthcoming water resource improvement attempts may employ this estimation approach for low flow forecasting, baseflow trend analysis, as well as planning and designing water resources projects.

Development of clustered polynomial chaos expansion model for stochastic hydrological prediction

This study introduced a clustered polynomial chaos expansion (CPCE) model to reveal random propagation and dynamic sensitivity of uncertainty parameters in hydrologic prediction. In the CPCE model, the random characteristics of the streamflow simulations resulting from parameter uncertainties are characterized through the polynomial chaos expansion (PCE) model based on the probabilistic collocation method. At the same time, a multivariate discrete non-functional relationship between PCE coefficients and hydrological model inputs is established based on stepwise cluster analysis. Therefore, compared with traditional PCE method, the developed CPCE model cannot only reflect uncertainty propagation in stochastic hydrological simulation, but also have the capability of random forecasting. Moreover, the dynamic sensitivities of model parameters are investigated through the multilevel factorial analyses. The developed approach was applied for streamflow forecasting for the Ruihe watershed, China. Results showed that with effective quantification for the random characteristics of hydrological processes, the CPCE model can directly predict runoff series and generate the associated probability distributions at different time periods. The dynamic sensitivity analysis indicates that the maximum soil moisture capacity within the catchment plays a key role in the accuracy of the low-flow forecasting, while the degree of spatial variability in soil moisture capacities has a remarkable impact on the accuracy of the high-flow forecasting in the studied watershed.

PREMHYCE: An operational tool for low-flow forecasting

Abstract. In many countries, rivers are the primary supply of water. A number of uses are concerned (drinking water, irrigation, hydropower, etc.) and they can be strongly affected by water shortages. Therefore, there is a need for the early anticipation of low-flow periods to improve water management. This is strengthened by the perspective of having more severe summer low flows in the context of climate change. Several French institutions (Inrae, BRGM, Météo-France, EDF and Lorraine University) have been collaborating over the last years to develop an operational tool for low-flow forecasting, called PREMHYCE. It was tested in real time on 70 catchments in continental France in 2017, and on 48 additional catchments in 2018. PREMHYCE includes five hydrological models: one uncalibrated physically-based model and four storage-type models of various complexity, which are calibrated on gauged catchments. The models assimilate flow observations or implement post-processing techniques. Low-flow forecasts can be issued up to 90 d ahead, based on ensemble streamflow prediction (ESP) using historical climatic data as ensembles of future input scenarios. These climatic data (precipitation, potential evapotranspiration and temperature) are provided by Météo-France with the daily gridded SAFRAN reanalysis over the 1958–2017 period, which includes a wide range of conditions. The tool provides numerical and graphical outputs, including the forecasted ranges of low flows, and the probability to be under low-flow warning thresholds provided by the users. Outputs from the different hydrological models can be combined through a simple multi-model approach to improve the robustness of forecasts. Results are illustrated for the Ill River at Didenheim (northeastern France) where the 2017 low-flow period was particularly severe and for which PREMHYCE provided useful forecasts.

Long short-term memory (LSTM) recurrent neural network for low-flow hydrological time series forecasting

This article explores the suitability of a long short-term memory recurrent neural network (LSTM-RNN) and artificial intelligence (AI) method for low-flow time series forecasting. The long short-term memory works on the sequential framework which considers all of the predecessor data. This forecasting method used daily discharged data collected from the Basantapur gauging station located on the Mahanadi River basin, India. Different metrics [root-mean-square error (RMSE), Nash–Sutcliffe efficiency (ENS), correlation coefficient (R) and mean absolute error] were selected to assess the performance of the model. Additionally, recurrent neural network (RNN) model is also used to compare the adaptability of LSTM-RNN over RNN and naive method. The results conclude that the LSTM-RNN model (R = 0.943, ENS = 0.878, RMSE = 0.487) outperformed RNN model (R = 0.935, ENS = 0.843, RMSE = 0.516) and naive method (R = 0.866, ENS = 0.704, RMSE = 0.793). The finding of this research concludes that LSTM-RNN can be used as new reliable AI technique for low-flow forecasting.

Seasonal streamflow forecasting by conditioning climatology with precipitation indices

Abstract. Many fields, such as drought-risk assessment or reservoir management, can benefit from long-range streamflow forecasts. Climatology has long been used in long-range streamflow forecasting. Conditioning methods have been proposed to select or weight relevant historical time series from climatology. They are often based on general circulation model (GCM) outputs that are specific to the forecast date due to the initialisation of GCMs on current conditions. This study investigates the impact of conditioning methods on the performance of seasonal streamflow forecasts. Four conditioning statistics based on seasonal forecasts of cumulative precipitation and the standardised precipitation index were used to select relevant traces within historical streamflows and precipitation respectively. This resulted in eight conditioned streamflow forecast scenarios. These scenarios were compared to the climatology of historical streamflows, the ensemble streamflow prediction approach and the streamflow forecasts obtained from ECMWF System 4 precipitation forecasts. The impact of conditioning was assessed in terms of forecast sharpness (spread), reliability, overall performance and low-flow event detection. Results showed that conditioning past observations on seasonal precipitation indices generally improves forecast sharpness, but may reduce reliability, with respect to climatology. Conversely, conditioned ensembles were more reliable but less sharp than streamflow forecasts derived from System 4 precipitation. Forecast attributes from conditioned and unconditioned ensembles are illustrated for a case of drought-risk forecasting: the 2003 drought in France. In the case of low-flow forecasting, conditioning results in ensembles that can better assess weekly deficit volumes and durations over a wider range of lead times.

LOW FLOW FORECASTING WITH RECESSION ANALYSIS APPROACHES

LOW FLOW FORECASTING WITH RECESSION ANALYSIS APPROACHES

Seasonality of low flows and dominant processes in the Rhine River

Low flow forecasting is crucial for sustainable cooling water supply and planning of river navigation in the Rhine River. The first step in reliable low flow forecasting is to understand the characteristics of low flow. In this study, several methods are applied to understand the low flow characteristics of Rhine River basin. In 108 catchments of the Rhine River, winter and summer low flow regions are determined with the seasonality ratio (SR) index. To understand whether different numbers of processes are acting in generating different low flow regimes in seven major sub-basins (namely, East Alpine, West Alpine, Middle Rhine, Neckar, Main, Mosel and Lower Rhine) aggregated from the 108 catchments, the dominant variable concept is adopted from chaos theory. The number of dominant processes within the seven major sub-basins is determined with the correlation dimension analysis. Results of the correlation dimension analysis show that the minimum and maximum required number of variables to represent the low flow dynamics of the seven major sub-basins, except the Middle Rhine and Mosel, is 4 and 9, respectively. For the Mosel and Middle Rhine, the required minimum number of variables is 2 and 6, and the maximum number of variables is 5 and 13, respectively. These results show that the low flow processes of the major sub-basins of the Rhine could be considered as non-stochastic or chaotic processes. To confirm this conclusion, the rescaled range analysis is applied to verify persistency (i.e. non-randomness) in the processes. The estimated rescaled range statistics (i.e. Hurst exponents) are all above 0.5, indicating that persistent long-term memory characteristics exist in the runoff processes. Finally, the mean values of SR indices are compared with the nonlinear analyses results to find significant relationships. The results show that the minimum and maximum numbers of required variables (i.e. processes) to model the dynamic characteristics for five out of the seven major sub-basins are the same, but the observed low flow regimes are different (winter low flow regime and summer low flow regime). These results support the conclusion that a few interrelated nonlinear variables could yield completely different behaviour (i.e. dominant low flow regime).

Application of radial basis function neural networks to short-term streamflow forecasting

The availability of adequate data is one of several factors that inform the choice of model used in various aspects of water resources research and management. Unfortunately, a number of models developed for use in water resources management find limited application due to their sophisticated data requirements. For a number of catchments in South Africa, the data available is insufficient for the adequate performance of some of these models. Consequently, for such catchments the ideal modeling approach would be one where the data available determines model requirements as opposed to the more conventional approach where the model data requirements are not readily available. For this reason, the use of radial basis function artificial neural networks in performing 1-day forecasts of streamflow was demonstrated in this study. Adept at simulating complex non-linear relationships, their consideration for use in streamflow modeling enabled the forecasting of streamflow in the Luvuvhu River in South Africa based on the data that was readily available. The suitability of the artificial neural network as a tool for streamflow forecasting was assessed by using the Nash–Sutcliffe efficiency (NSE) and the percent bias (% Bias) measures between the simulated and observed data. Over two of the reaches on the river, satisfactory model performance was obtained with the average NSE and % Bias measures of 0.988 and −22.76% respectively achieved during the calibration phase and 0.980 and −20.30% during the verification phase. This study also demonstrated that the RBFN are amenable, through choice of objective function, to forecast with greater accuracy portions of the hydrograph such as the situation may warrant for instance when low flow forecasting is of significance. The results suggest that artificial neural networks hold promise for streamflow forecasting in South Africa.

Book Reviews

Book Reviews

Development of a low-flow forecasting model using the M5 machine learning method

By using the M5 machine learning method and analysis of the recorded recession streamflow data, we modelled the streamflow recession coefficient k as a function of the flow rate at which the 7-day low flow forecast is made and the decrease in the flow rate from the previous day. Low flow forecasting models for seven of the Slovenian tributaries of the Sava River were developed and verification of the results showed improved accuracy of the models compared to when using a single-valued recession coefficient. From the structure of the models, we can also learn that the degree of the flow rate change in the last 24 hours can tell us more about the streamflow recession dynamics in the next seven days than the flow rate at the time when the forecast is made.

Low-Flow Forecasting on the Delaware River

A method is described for the design of reservoir releases to augment low flows of Delaware River, which was in the center of intensity of the 1961-1966 drought in the Northeastern United States. Index stream gages, weather forecasts, and anticipated hydroelectric plant operations are used to arrive at adequate reservoir release requirements. Low-flow augmentation to maintain certain basic rates of streamflow was provided for under the terms of a decree of the U. S. Supreme Court. A period of low flow on the Delaware River is expected every year.