Comparison of data-driven modelling techniques for river flow forecasting

Abstract

Accurate forecasting of streamflow is essential for the efficient operation of water resources systems. The streamflow process is complex and highly nonlinear. Therefore, researchers try to devise alterative techniques to forecast streamflow with relative ease and reasonable accuracy, although traditional deterministic and conceptual models are available. The present work uses three data-driven techniques, namely artificial neural networks (ANN), genetic programming (GP) and model trees (MT) to forecast river flow one day in advance at two stations in the Narmada catchment of India, and the results are compared. All the models performed reasonably well as far as accuracy of prediction is concerned. It was found that the ANN and MT techniques performed almost equally well, but GP performed better than both these techniques, although only marginally in terms of prediction accuracy in normal and extreme events.

Citation Londhe, S. & Charhate, S. (2010) Comparison of data-driven modelling techniques for river flow forecasting. Hydrol. Sci. J. 55(7), 1163–1174.     

Improving the forecasts of extreme streamflow by support vector regression with the data extracted by self‐organizing map

During typhoons or storms, accurate forecasts of hourly streamflow are necessary for flood warning and mitigation. However, hourly streamflow is difficult to forecast because of the complex physical process and the high variability in time. Furthermore, under the global warming scenario, events with extreme streamflow may occur that leads to more difficulties in forecasting streamflows. Hence, to obtain more accurate hourly streamflow forecasts, an improved streamflow forecasting model is proposed in this paper. The computational kernel of the proposed model is developed on the basis of support vector machine (SVM). Additionally, self‐organizing map (SOM) is used to analyse observed data to extract data with specific properties, which are capable of providing valuable information for streamflow forecasting. After reprocessing, these extracted data and the observed data are used to construct the SVM‐based model. An application is conducted to clearly demonstrate the advantage of the proposed model. The comparison between the proposed model and the conventional SVM model, which is constructed without SOM, is performed. The results indicate that the proposed model is better performed than the conventional SVM model. Moreover, as regards the extreme events, the result shows that the proposed model reduces the forecasting error, especially the error of peak streamflow. It is confirmed that because of the use of data extracted by SOM, the improved forecasting performance is obtained. The proposed model, which can produce accurate forecasts, is expected to be useful to support flood warning systems. Copyright © 2012 John Wiley & Sons, Ltd.     

Wavelet regression model as an alternative to neural networks for monthly streamflow forecasting

The accuracy of the wavelet regression (WR) model in monthly streamflow forecasting is investigated in the study. The WR model is improved combining the two methods—the discrete wavelet transform (DWT) model and the linear regression (LR) model—for 1‐month‐ahead streamflow forecasting. In the first part of the study, the results of the WR model are compared with those of the single LR model. Monthly flow data from two stations, Gerdelli Station on Canakdere River and Isakoy Station on Goksudere River, in Eastern Black Sea region of Turkey are used in the study. The comparison results reveal that the WR model could increase the forecast accuracy of the LR model. In the second part of the study, the accuracy of the WR model is compared with those of the artificial neural networks (ANN) and auto‐regressive (AR) models. On the basis of the results, the WR is found to be better than the ANN and AR models in monthly streamflow forecasting. Copyright © 2009 John Wiley & Sons, Ltd.     

Sequential assimilation of soil moisture and streamflow data in a conceptual rainfall–runoff model

Soil moisture is a key hydrological variable in flood forecasting: it largely influences the partition of rain between runoff and infiltration and thus controls the flow at the outlet of a catchment. The methodology developed in this paper aims at improving the commonly used hydrological tools in an operational forecasting context by introducing soil moisture data into streamflow modelling. A sequential assimilation procedure, based on an extended Kalman filter, is developed and coupled with a lumped conceptual rainfall–runoff model. It updates the internal states of the model (soil and routing reservoirs) by assimilating daily soil moisture and streamflow data in order to better fit these external observations. We present in this paper the results obtained on the Serein, a Seine sub-catchment (France), during a period of about 2 years and using Time Domain Reflectivity probe soil moisture measurements from 0–10 to 0–100 cm and stream gauged data. Streamflow prediction is improved by assimilation of both soil moisture and streamflow individually and by coupled assimilation. Assimilation of soil moisture data is particularly effective during flood events while assimilation of streamflow data is more effective for low flows. Combined assimilation is therefore more adequate on the entire forecasting period. Finally, we discuss the adequacy of this methodology coupled with Remote Sensing data.     

Incorporating the spatio-temporal distribution of rainfall and basin geomorphology into nonlinear analyses of streamflow dynamics

Many recent studies have been devoted to the investigation of the nonlinear dynamics of rainfall or streamflow series based on methods of dynamical systems theory. Although finding evidence for the existence of a low-dimensional deterministic component in rainfall or streamflow is of much interest, not much attention has been given to the nonlinear dependencies of the two and especially on how the spatio-temporal distribution of rainfall affects the nonlinear dynamics of streamflow at flood time scales. In this paper, a methodology is presented which simultaneously considers streamflow series, spatio-temporal structure of precipitation and catchment geomorphology into a nonlinear analysis of streamflow dynamics. The proposed framework is based on “hydrologically-relevant” rainfall-runoff phase-space reconstruction acknowledging the fact that rainfall-runoff is a stochastic spatially extended system rather than a deterministic multivariate one. The methodology is applied to two basins in Central North America using 6-hour streamflow data and radar images for a period of 5 years. The proposed methodology is used to: (a) quantify the nonlinear dependencies between streamflow dynamics and the spatio-temporal dynamics of precipitation; (b) study how streamflow predictability is affected by the trade-offs between the level of detail necessary to explain the spatial variability of rainfall and the reduction of complexity due to the smoothing effect of the basin; and (c) explore the possibility of incorporating process-specific information (in terms of catchment geomorphology and an a priori chosen uncertainty model) into nonlinear prediction. Preliminary results are encouraging and indicate the potential of using the proposed methodology to understand via nonlinear analysis of observations (i.e., not based on a particular rainfall-runoff model) streamflow predictability and limits to prediction as a function of the complexity of spatio-temporal forcing relative to basin geomorphology.     

Determination of runoff components using path analysis and isotopic measurements in a glacier‐covered alpine catchment (upper Hailuogou Valley) in southwest China

Monitoring of stable water isotopes (δ¹⁸O and δ²H) at the watershed scales can improve our understanding of complex hydrology and hydroclimatology of the watershed, especially in remote regions. Previous studies that used tracers for hydrograph separation are largely based on end‐member mixing approach (EMMA), but one drawback of this approach is that at least two independent tracers are required for multi‐component separation. Here we introduce a new approach—path analysis, in combination with isotopic measurements to investigate the runoff generation in a glacier‐covered alpine catchment (upper Hailuogou Valley) in southwest China. This newly developed method can not only provide a multi‐component hydrograph separation with the aid of only one tracer but also determine the direct and indirect influence of sources on streamflow. Path analysis show that the majority of streamflow is dominated by ice/snow meltwater that represents about 63–78% of the total discharge, whereas precipitation and groundwater contribute approximately 19–39% and 2–4% of the streamflow discharge, respectively. These results are in good agreement with those derived from EMMA (using ¹⁸O and Cl^? as tracers), corroborating that our proposed approach is successful in hydrograph separation of the catchment. This approach may provide new opportunities for the hydrograph separation of catchment with sparse data and be of interest to catchment hydrologists who seek to understand the behaviour of hydrologic systems. Copyright © 2015 John Wiley & Sons, Ltd.     

Evaluation of ocean-atmospheric indices as predictors for summer streamflow of the Yangtze River based on ROC analysis

Antecedent anomalies of sea surface temperature and atmospheric circulation are important signals for making long-term streamflow forecasts. In this study, four groups of ocean-atmospheric indices, i.e, El Niño Southern Oscillation (ENSO), the Northern Hemisphere atmospheric circulation, the Southern Hemisphere atmospheric circulation (SAC), and the Western Pacific and Indian Ocean SST (WPI), are evaluated for forecasting summer streamflow of the Yangtze River. The gradient boosting regression tree (GBRT) is used to forecast streamflow based on each group of indices. The score based on receiver operating characteristics (ROC) curves, i.e., area under the ROC curve (AUC), is used to evaluate skills of models for identifying the high category and the low category of summer streamflow. It is found that the ENSO group and the SAC group show higher AUC values. Furthermore, both AUC values of GBRT models and individual indices show that the low flow years are easier to be identified than the high flow years. The result of this study highlights the skill from the Southern Hemisphere circulation systems for forecasting summer streamflow of the Yangtze River. Results of relative influences of predictors in GBRT models and AUC of individual indices indicate some key ocean-atmospheric indices, such as the Multivariate ENSO Index and the 500-hPa height of the east of Australia.     

Inferring catchment flow path responses using a data-driven model: an exploratory study based on a generalized additive model

Identifying physical catchment processes from streamflow data, such as quick- and slow-flow paths, remains challenging. This study is designed to explore whether a flexible nonparametric regression model (generalized additive model, GAM) can be used to infer different flow paths. This assumes that the data relationship in data-driven models is also a reflection of catchment physical processes. The GAM, using time-lagged flow covariates, was fitted to synthetic rainfall–runoff data simulated using simple linear reservoirs. Partial plots of the time-lagged covariates show that the model could differentiate simple and more complex flow paths in simulated synthetic data with short and long memory systems and varying between dry and wet climates. Further analysis of data from real catchments showed that the model could differentiate catchments dominated by slow flow and by quick flow. Therefore, this study indicates that GAM can be used to identify catchment storages and delay processes from streamflow data.     

Assessing the impact of climate variability and human activities on streamflow from the Wuding River basin in China

Located in the Loess Plateau of China, the Wuding River basin (30 261 km²) contributes significantly to the total sediment yield in the Yellow River. To reduce sediment yield from the catchment, large-scale soil conservation measures have been implemented in the last four decades. These included building terraces and sediment-trapping dams and changing land cover by planting trees and improving pastures. It is important to assess the impact of these measures on the hydrology of the catchment and to provide a scientific basis for future soil conservation planning. The non-parametric Mann–Kendall–Sneyers rank test was employed to detect trends and changes in annual streamflow for the period of 1961 to 1997. Two methods were used to assess the impact of climate variability on mean annual streamflow. The first is based on a framework describing the sensitivity of annual streamflow to precipitation and potential evaporation, and the second relies on relationships between annual streamflow and precipitation. The two methods produced consistent results. A significant downward trend was found for annual streamflow, and an abrupt change occurred in 1972. The reduction in annual streamflow between 1972 and 1997 was 42% compared with the baseline period (1961–1971). Flood-season streamflow showed an even greater reduction of 49%. The streamflow regime of the catchment showed a relative reduction of 31% for most percentile flows, except for low flows, which showed a 57% reduction. The soil conservation measures reduced streamflow variability, leading to more uniform streamflow. It was estimated that the soil conservation measures account for 87% of the total reduction in mean annual streamflow in the period of 1972 to 1997, and the reduction due to changes in precipitation and potential evaporation was 13%. Copyright © 2007 John Wiley & Sons, Ltd.     

Effects of precipitation and landuse on runoff during the past 50 years in a typical watershed in the Loess Plateau, China

In the past century, great progress has been made worldwide in our understanding of forest-water relationship. The successful forestation programs implemented in China-which have improved the ecological environmental conditions-have gained the attention of many researchers and highlighted the relationship between forestation and water yields. The arid and semi-arid Loess Plateau has received attention from water engineers and eco-hydrological researchers in China because of a shortage in water resources. We selected one of the oldest stations conducting soil and water conservation experiments, the Xifeng soil and water conservation station, and chose the Nanxiaohe catchment and its paired catchments （Yangjiagou catchment and Dongzhuanggou catchment） as our research areas. Trends in precipitation, air temperature, streamflow over the past 50 years, and the effect of changing land use on streamflow were analyzed. The Mann-Kendall test showed that precipitation had a negative trend （downward trend）, whereas air temperature showed a positive trend （upward trend） from the past to present in the Nanxiaohe catchment. However, the trends seen in precipitation, air temperature did not contain any ＂jumping points.＂ The paired catchment approach is used to detect the effects of land cover change on hydrology in the Yangjiagou and the contrast catchment, i.e., Dongzhuanggou catchment in our study. The results showed a large change in land use in the Yangjiagou catchment from 1954 to 2008. An increase in forested land （from 0% to 40.08% from 1954 to 2008） and a reduction of bare land （from 51.26% to 5.50% from 1954 to 2008） accounted for a large part of the change in land use. However, the land use changed little in the contrast catchment. The comparison of streamfiow in the paired catchments showed that forestation reduced streamflow by 49.63% （or 6.5 mm） each year.     

A univariate model for long-term streamflow forecasting

This paper, the first in a series of two, employs the principle of maximum entropy (POME) via maximum entropy spectral analysis (MESA) to develop a univariate model for long-term streamflow forecasting. Three cases of streamflow forecasting are investigated: forward forecasting, backward forecasting (or reconstruction) and intermittent forecasting (or filling in missing records). Application of the model is discussed in the second paper.     

RELATIONSHIPS BETWEEN CATCHMENT ATTRIBUTES AND HYDROLOGICAL RESPONSE CHARACTERISTICS IN SMALL AUSTRALIAN MOUNTAIN ASH CATCHMENTS

Sixteen small catchments in the Maroondah region of Victoria, Australia were analysed using rainfall, temperature and streamflow time series with a rainfall–runoff model whose parameters efficiently characterize the hydrological response of a catchment. A set of catchment attributes for each of these catchments was then compared with the associated set of hydrological response characteristics of the catchments as estimated by the model. The time constant governing quickflow recession of streamflow (τ_q) was related to the drainage network and catchment area. The time constant governing slowflow recession of streamflow (τ_s) was related to the slope and shape of the catchment. The parameter governing evapotranspirative losses ( f ) was related to catchment gradient and vegetative water use. Forestry activities in the catchments changed evapotranspirative losses and thus total volume of streamflow, but did not affect the rate of streamflow recession.     

Streamflow drought time series forecasting

Drought is considered to be an extreme climatic event causing significant damage both in the natural environment and in human lives. Due to the important role of drought forecasting in water resources planning and management and the stochastic behavior of drought, a multiplicative seasonal autoregressive integrated moving average (SARIMA) model is applied to the monthly streamflow forecasting of the Zayandehrud River in western Isfahan province, Iran. After forecasting 12 leading month streamflow, four drought thresholds including streamflow mean, monthly streamflow mean, 2-, 5-, 10- and 20-year return period monthly drought and standardized streamflow index were chosen. Both observed and forecasted streamflow showed a drought period with different severity in the lead-time. This study also demonstrates the usefulness of SARIMA models in forecasting, water resources planning and management.     

Alteration of hydrological processes and streamflow with juniper (Juniperus virginiana) encroachment in a mesic grassland catchment

To understand the effect of woody plant encroachment on hydrological processes of mesic grasslands, we quantified infiltration capacity in situ, the temporal changes in soil water storage, and streamflow of a grassland catchment and a catchment heavily encroached by juniper (Juniperus virginiana, eastern redcedar) in previously cultivated, non‐karst substrate grasslands in north‐central Oklahoma for 3 years. The initial and steady‐state infiltration rates under the juniper canopy were nearly triple to that of the grassland catchment and were intermediate in the intercanopy spaces within the encroached catchment. Soil water content and soil water storage on the encroached catchment were generally lower than on the grassland catchment, especially when preceding the seasons of peak rainfall in spring and fall. Frequency and magnitude of streamflow events were reduced in the encroached catchment. Annual runoff coefficients for the encroached catchment averaged 2.1%, in contrast to 10.6% for the grassland catchment. Annual streamflow duration ranged from 80 to 250 h for the encroached catchment compared with 600 to 800 h for the grassland catchment. Our results showed that the encroachment of juniper into previously cultivated mesic grasslands fundamentally alters catchment hydrological function. Rapid transformation of mesic grassland to a woodland state with juniper encroachment, if not confined, has the potential to drastically reduce soil water, streamflow and flow duration of ephemeral streams in the Southern Great Plains. Copyright © 2013 John Wiley & Sons, Ltd.     

Can climate variability information constrain a hydrological model for an ungauged Costa Rican catchment?

Long‐term hydrological data are key to understanding catchment behaviour and for decision making within water management and planning. Given the lack of observed data in many regions worldwide, such as Central America, hydrological models are an alternative for reproducing historical streamflow series. Additional types of information—to locally observed discharge—can be used to constrain model parameter uncertainty for ungauged catchments. Given the strong influence that climatic large‐scale processes exert on streamflow variability in the Central American region, we explored the use of climate variability knowledge as process constraints to constrain the simulated discharge uncertainty for a Costa Rican catchment, assumed to be ungauged. To reduce model uncertainty, we first rejected parameter relationships that disagreed with our understanding of the system. Then, based on this reduced parameter space, we applied the climate‐based process constraints at long‐term, inter‐annual, and intra‐annual timescales. In the first step, we reduced the initial number of parameters by 52%, and then, we further reduced the number of parameters by 3% with the climate constraints. Finally, we compared the climate‐based constraints with a constraint based on global maps of low‐flow statistics. This latter constraint proved to be more restrictive than those based on climate variability (further reducing the number of parameters by 66% compared with 3%). Even so, the climate‐based constraints rejected inconsistent model simulations that were not rejected by the low‐flow statistics constraint. When taken all together, the constraints produced constrained simulation uncertainty bands, and the median simulated discharge followed the observed time series to a similar level as an optimized model. All the constraints were found useful in constraining model uncertainty for an—assumed to be—ungauged basin. This shows that our method is promising for modelling long‐term flow data for ungauged catchments on the Pacific side of Central America and that similar methods can be developed for ungauged basins in other regions where climate variability exerts a strong control on streamflow variability.     

Estimation of actual evapotranspiration from an alluvial aquifer of the Kouris catchment (Cyprus) using continuous streamflow records

The Kouris catchment is located in the south of the Troodos massif in Cyprus. The hydrology is driven by a Mediterranean climate, a mountainous topography, and a complex distribution of hydrogeological properties resulting from complex geology. To quantify the regional water balance further, a simple method using continuous streamflow records in the River Limnatis (Kouris catchment) was applied to calculate the actual evapotranspiration rate in the dry seasons. It was found that daily cycles of streamflow, recorded by automatic pressure logger, were caused by direct evaporation from the groundwater table and by transpiration of riparian forest. The daily amounts of ‘missing’ streamflow were calculated for the period 30 October–4 November 2001 and were extrapolated to the entire dry season and to the whole Kouris catchment. The actual evapotranspiration rate from the alluvial aquifer of the region is 2·4 ± 0·5 Mm³ for April–September 2001. The validity of the assumptions and the uncertainties in the estimates used in the method are discussed. Copyright © 2005 John Wiley & Sons, Ltd.     

Constraining water age dynamics in a south-eastern Australian catchment using an age-ranked storage and stable isotope approach

Improving our knowledge of the travel times of water through catchments is critical for the management and protection of water resources and to improve our understanding of fundamental catchment behaviour. In this study we use the age-ranked storage framework StorAge Selection (SAS) to investigate travel times in the Corin catchment, a headwater catchment in the south-east of Australia covered by native Eucalyptus species. Few studies have applied the SAS framework globally and in energy-intensive areas where catchment losses are heavily in favour of evapotranspiration relative to streamflow. A combination of observed and modelled values of oxygen-18 (δ¹⁸O ), the stable isotope in water, are used to constrain storage selection preferences of streamflow and evapotranspiration and the size of the catchment active storage. The highest performing parameter combinations that could reproduce δ¹⁸O in streamflow were dependent on a strong preference for young water in evapotranspiration, and a mixture of weak young and old water preference in streamflow. The mean travel time of streamflow over the study period 2007–2019, weighted by the flow rate, is limited to within a probable range of 2.81–9.77 years. The size of the active storage, a key parameter in the SAS framework, was poorly identified, and in combination with the isotopic inputs into the model, contributed to the uncertainty of the results. We discuss the implications of the results with respect to the study area, as well as within the context of SAS research globally and identify ways to improve the modelling process.     

A comparison of streamflow,salt and water balances in adjacent farmland and forest catchments in south‐western Victoria,Australia

A study of the hydrologic effects of catchment change from pasture to plantation was carried out in Gatum, south‐western Victoria, Australia. This study describes the hydrologic characteristics of two adjacent catchments: one with 97% grassland and the other one with 62% Eucalyptus globulus plantations. Streamflow from both catchments was intermittent during the 20‐month study period. Monthly streamflow was always greater in the pasture‐dominated catchment compared with the plantation catchment because of lower evapotranspiration in the pasture‐based catchment. This difference in streamflow was also observed even during summer 2010/2011 when precipitation was 74% above average (1954–2012) summer rainfall. Streamflow peaks in the plantation‐based catchment were smaller than in the pasture‐dominated system. Flow duration curves show differences between the pasture and plantation‐dominated catchments and affect both high‐flow and low‐flow periods. Groundwater levels fell (up to 4.4 m) in the plantation catchment during the study period but rose (up to 3.2 m) in the pasture catchment. Higher evapotranspiration in the plantation catchment resulted in falling groundwater levels and greater disconnection of the groundwater system from the stream, resulting in lower baseflow contribution to streamflow. Salt export from each catchment increases with increasing flow and is higher at the pasture catchment, mainly because of the higher flow. Reduced salt loading to streams due to tree planting is generally considered environmentally beneficial in saline areas of south‐eastern Australia, but this benefit is offset by reduced total streamflow. Copyright © 2014 John Wiley & Sons, Ltd.     

A comprehensive assessment framework for attributing trends in streamflow and groundwater storage to climatic and anthropogenic changes: A case study in the typical semi-arid catchments of southern India

The clearest signs of hydrologic change can be observed from the trends in streamflow and groundwater levels in a catchment. During 1980–2007, significant declines in streamflow (−3.03 mm/year) and groundwater levels (−0.22 m/year) were observed in Himayat Sagar (HS) catchment, India. We examined the degree to which hydrologic changes observed in the HS catchment can be attributed to various internal and external drivers of change (climatic and anthropogenic changes). This study used an investigative approach to attribute hydrologic changes. First, it involves to develop a model and test its ability to predict hydrologic trends in a catchment that has undergone significant changes. Second, it examines the relative importance of different causes of change on the hydrologic response. The analysis was carried out using Modified Soil and Water Assessment Tool (SWAT), a semi-distributed rainfall-runoff model coupled with a lumped groundwater model for each sub- catchment. The model results indicated that the decline in potential evapotranspiration (PET) appears to be partially offset by a significant response to changes in rainfall. Measures that enhance recharge, such as watershed hydrological structures, have had limited success in terms of reducing impacts on the catchment-scale water balance. Groundwater storage has declined at a rate of 5 mm/y due to impact of land use changes and this was replaced by a net addition of 2 mm/y by hydrological structures. The impact of land use change on streamflow is an order of magnitude larger than the impact of hydrological structures and about is 2.5 times higher in terms of groundwater impact. Model results indicate that both exogenous and endogenous changes can have large impacts on catchment hydrology and should be considered together. The proposed comprehensive framework and approach demonstrated here is valuable in attributing trends in streamflow and groundwater levels to catchment climatic and anthropogenic changes.