River Flow Forecasting Research Articles

Application of the SWAT Model for River Flow Forecasting in Sri Lanka

Application of the SWAT Model for River Flow Forecasting in Sri Lanka

Comparison of different forms of the Multi-layer Feed-Forward Neural Network method used for river flow forecasting

Abstract. The Multi-Layer Feed-Forward Neural Network (MLFFNN) is applied in the context of river flow forecast combination, where a number of rainfall-runoff models are used simultaneously to produce an overall combined river flow forecast. The operation of the MLFFNN depends not only on its neuron configuration but also on the choice of neuron transfer function adopted, which is non-linear for the hidden and output layers. These models, each having a different structure to simulate the perceived mechanisms of the runoff process, utilise the information carrying capacity of the model calibration data in different ways. Hence, in a discharge forecast combination procedure, the discharge forecasts of each model provide a source of information different from that of the other models used in the combination. In the present work, the significance of the choice of the transfer function type in the overall performance of the MLFFNN, when used in the river flow forecast combination context, is investigated critically. Five neuron transfer functions are used in this investigation, namely, the logistic function, the bipolar function, the hyperbolic tangent function, the arctan function and the scaled arctan function. The results indicate that the logistic function yields the best model forecast combination performance. Keywords: River flow forecast combination, multi-layer feed-forward neural network, neuron transfer functions, rainfall-runoff models

Evaluation of artificial neural network techniques for flow forecasting in the River Yangtze, China

Abstract. While engineers have been quantifying rainfall-runoff processes since the mid-19th century, it is only in the last decade that artificial neural network models have been applied to the same task. This paper evaluates two neural networks in this context: the popular multilayer perceptron (MLP), and the radial basis function network (RBF). Using six-hourly rainfall-runoff data for the River Yangtze at Yichang (upstream of the Three Gorges Dam) for the period 1991 to 1993, it is shown that both neural network types can simulate river flows beyond the range of the training set. In addition, an evaluation of alternative RBF transfer functions demonstrates that the popular Gaussian function, often used in RBF networks, is not necessarily the ‘best’ function to use for river flow forecasting. Comparisons are also made between these neural networks and conventional statistical techniques; stepwise multiple linear regression, auto regressive moving average models and a zero order forecasting approach. Keywords: Artificial neural network, multilayer perception, radial basis function, flood forecasting

Multi-model data fusion for river flow forecasting: an evaluation of six alternative methods based on two contrasting catchments

Abstract. This paper evaluates six published data fusion strategies for hydrological forecasting based on two contrasting catchments: the River Ouse and the Upper River Wye. The input level and discharge estimates for each river comprised a mixed set of single model forecasts. Data fusion was performed using: arithmetic-averaging, a probabilistic method in which the best model from the last time step is used to generate the current forecast, two different neural network operations and two different soft computing methodologies. The results from this investigation are compared and contrasted using statistical and graphical evaluation. Each location demonstrated several options and potential advantages for using data fusion tools to construct superior estimates of hydrological forecast. Fusion operations were better in overall terms in comparison to their individual modelling counterparts and two clear winners emerged. Indeed, the six different mechanisms on test revealed unequal aptitudes for fixing different categories of problematic catchment behaviour and, in such cases, the best method(s) were a good deal better than their closest rival(s). Neural network fusion of differenced data provided the best solution for a stable regime (with neural network fusion of original data being somewhat similar) — whereas a fuzzified probabilistic mechanism produced a superior output in a more volatile environment. The need for a data fusion research agenda within the hydrological sciences is discussed and some initial suggestions are presented. Keywords: data fusion, fuzzy logic, neural network, hydrological modelling

River flow forecasting: use of phase-space reconstruction and artificial neural networks approaches

The use of two non-linear black-box approaches, phase-space reconstruction (PSR) and artificial neural networks (ANN), for forecasting river flow dynamics is studied and a comparison of their performances is made. This is done by attempting 1-day and 7-day ahead forecasts of the daily river flow from the Nakhon Sawan station at the Chao Phraya River basin in Thailand. The results indicate a reasonably good performance of both approaches for both 1-day and 7-day ahead forecasts. However, the performance of the PSR approach is found to be consistently better than that of ANN. One reason for this could be that in the PSR approach the flow series in the phase-space is represented step by step in local neighborhoods, rather than a global approximation as is done in ANN. Another reason could be the use of the multi-layer perceptron (MLP) in ANN, since MLPs may not be most appropriate for forecasting at longer lead times. The selection of training set for the ANN may also contribute to such results. A comparison of the optimal number of variables for capturing the flow dynamics, as identified by the two approaches, indicates a large discrepancy in the case of 7-day ahead forecasts (1 and 7 variables, respectively), though for 1-day ahead forecasts it is found to be consistent (3 variables). A possible explanation for this could be the influence of noise in the data, an observation also made from the 1-day ahead forecast results using the PSR approach. The present results lead to observation on: (1) the use of other neural networks for runoff forecasting, particularly at longer lead times; (2) the influence of training set used in the ANN; and (3) the effect of noise on forecast accuracy, particularly in the PSR approach.

Using a lognormal diffusion process to forecast river flows

The logarithm of the flow of a river in Quebec, Canada, is modeled as a Gaussian diffusion process. Using this model, forecasts of the river flow are made for up to 7 days ahead. The forecasts are found to be much more accurate for 1 day ahead than those produced by a deterministic model called PREVIS. The stochastic model also outperforms PREVIS for 2 days ahead forecasts and is comparable to PREVIS for 3 days ahead forecasts. In general, PREVIS is more accurate than the stochastic model proposed from 4 days ahead forecasts. Various possible improvements of the stochastic model are considered, including the incorporation of exogenous variables and the use of linear regression. The best forecasts are actually obtained by taking the average value of the forecasts produced by the stochastic model and by PREVIS. Finally, confidence interval for the forecasted values are also provided.

A non-linear neural network technique for updating of river flow forecasts

Abstract. A non-linear Auto-Regressive Exogenous-input model (NARXM) river flow forecasting output-updating procedure is presented. This updating procedure is based on the structure of a multi-layer neural network. The NARXM-neural network updating procedure is tested using the daily discharge forecasts of the soil moisture accounting and routing (SMAR) conceptual model operating on five catchments having different climatic conditions. The performance of the NARXM-neural network updating procedure is compared with that of the linear Auto-Regressive Exogenous-input (ARXM) model updating procedure, the latter being a generalisation of the widely used Auto-Regressive (AR) model forecast error updating procedure. The results of the comparison indicate that the NARXM procedure performs better than the ARXM procedure. Keywords: Auto-Regressive Exogenous-input model, neural network, output-updating procedure, soil moisture accounting and routing (SMAR) model

Multivariate nonlinear prediction of river flows

This paper considers a nonlinear method for the forecasting of river flows, developed in its univariate form in the context of modern nonlinear time series analysis by Farmer and Sidorowich [Phys. Rev. Lett. 59 (1987) 845] and usually referred to as nonlinear prediction (NLP). Here, such a method is extended to a multivariate form to include information from other time series in addition to that of discharge. The conceptual basis of the multivariate approach is explained in the first part of the paper, while the second part deals with the application to forecasting of river flow. The good quality of the predictions obtained, along with the flexibility of the multivariate method to adapt to the different sources of information, indicates that this technique could be of real interest in the hydrologic field.

Seasonal Forecasting of River Flows in the British Isles Using North Atlantic Pressure Patterns

AbstractA potentially useful forecasting relationship is demonstrated between the North Atlantic Oscillation index in winter (January‐February) and the following summer (July‐September) monthly mean flows for selected rivers in the British Isles. The relationship was strongest in August when up to 40% of the variance in monthly mean flow may be explained. For two rivers in southern and eastern England, positive phases of the winter North Atlantic Oscillation index were found to precede summer flows which were nearly 50% of long‐term average. A regression equation which was established to predict August flows in the Great Stour produced a correlation score of 0.6 at a lead time of six months. Further research is needed to determine the significance of catchment characteristics and geographic location relative to forecasting skill.

River flow forecasting using artificial neural networks

Abstract River flow forecasting is required to provide basic information on a wide range of problems related to the design and operation of river systems. The availability of extended records of rainfall and other climatic data, which could be used to obtain stream flow data, initiated the practice of rainfall-runoff modelling. While conceptual or physically-based models are of importance in the understanding of hydrological processes, there are many practical situations where the main concern is with making accurate predictions at specific locations. In such situation it is preferred to implement a simple “black box” (data-driven, or machine learning) model to identify a direct mapping between the inputs and outputs without detailed consideration of the internal structure of the physical process. Artificial neural networks (ANNs) is probably the most successful machine learning technique with flexible mathematical structure which is capable of identifying complex non-linear relationships between input and output data without attempting to reach understanding as to the nature of the phenomena. In this study the applicability of ANNs for downstream flow forecasting in the Apure river basin (Venezuela) was investigated. Two types of ANN architectures, namely multi-layer perceptron network (MLP) and a radial basis function network (RBF) were implemented. The performances of these networks were compared with a conceptual rainfall-runoff model and they were found to be slightly better for this river flow-forecasting problem.

Inductive learning approaches to rainfall-runoff modelling.

Trying to model the rainfall-runoff process is a complex activity as it is influenced by a number of implicit and explicit factors--for example, precipitation distribution, evaporation, transpiration, abstraction, watershed topography, and soil types. However, this kind of forecasting is particularly important as it is used to predict serious flooding, estimate erosion and identify problems associated with low flow. Inductive learning approaches (e.g. decision trees and artificial neural networks) are particularly well suited to problems of this nature as they can often interpret underlying factors (such as seasonal variations) which cannot be modelled by other techniques. In addition, these approaches can easily be trained on the explicit factors (e.g. rainfall) and the inexplicit factors (e.g. abstraction) that affect river flow. Inductive learning approaches can also be extended to account for new factors that emerge over a period of time. This paper evaluates the application of decision trees and two artificial neural network models (the multilayer perceptron and the radial basis function network) to river flow forecasting in two flood prone UK catchments using real hydrometric data. Comparisons are made between the performance of these approaches and conventional flood forecasting systems.

Comparing neural network and autoregressive moving average techniques for the provision of continuous river flow forecasts in two contrasting catchments

The forecasting power of neural network (NN) and autoregressive moving average (ARMA) models are compared. Modelling experiments were based on a 3-year period of continuous river flow data for two contrasting catchments: the Upper River Wye and the River Ouse. Model performance was assessed using global and storm-specific quantitative evaluation procedures. The NN and ARMA solutions provided similar results, although naive predictions yielded poorer estimates. The annual data were then grouped into a set of distinct hydrological event types using a self-organizing map and two rising event clusters were modelled using the NN technique. These alternative investigations provided encouraging results. Copyright © 2000 John Wiley & Sons, Ltd.

A comparison of artificial neural networks used for river forecasting

Abstract. This paper compares the performance of two artificial neural network (ANN) models – the multi layer perceptron (MLP) and the radial basis function network (RBF) – with a stepwise multiple linear regression model (SWMLR) and zero order forecasts (ZOF) of river flow. All models were trained using 15 minute rainfall-runoff data for the River Mole, a flood-prone tributary of the River Thames, UK. The models were then used to forecast river flows with a 6 hour lead time and 15 minute resolution, given only antecedent rainfall and discharge measurements. Two seasons (winter and spring) were selected for model testing using a cross-validation technique and a range of diagnostic statistics. Overall, the MLP was more skillful than the RBF, SWMLR and ZOF models. However, the RBF flow forecasts were only marginally better than those of the simpler SWMLR and ZOF models. The results compare favourably with a review of previous studies and further endorse claims that ANNs are well suited to rainfall-runoff modelling and (potentially) real-time flood forecasting.

A real-time combination method for the outputs of different rainfall-runoff models

Application of the concept of combining the estimated forecast output of different rainfall-runoff models to yield an overall combined estimated output in the context of real-time river flow forecasting is explored. A Real-Time Model Output Combination Method (RTMOCM) is developed, based on the structure of the Linear Transfer Function Model (LTFM) and utilizing the concept of the Weighted Average Method (WAM) for model output combination. A multiple-input single-output form of the LTFM is utilized in the RTMOCM. This form of the LTFM model uses synchronously the daily simulation-mode model-estimated discharge time series of the rainfall-runoff models selected for combination, its inherent updating structure being used for providing updated combined discharge forecasts. The RTMOCM is applied to the daily data of five catchments, using the simulation-mode estimated discharges of three selected rainfall-runoff models, comprising one conceptual model (Soil Moisture Accounting and Routing Procedure—SMAR) and two black-box models (Linear Perturbation Model—LPM and Linearly-Varying Variable Gain Factor Model—LVGFM). In order to get an indication of the accuracy of the updated combined discharge forecasts relative to the updated discharge forecasts of the individual models, the LTFM is also used for updating the simulation-mode discharge time series of each of the three individual models. The results reveal that the updated combined discharge forecasts provided by the RTMOCM, with parameters obtained by linear regression, can improve on the updated discharge forecasts of the individual rainfall-runoff models.

Parsimonious modelling of water and suspended sediment flux from nested catchments affected by selective tropical forestry.

The ability to model the suspended sediment flux (SSflux) and associated water flow from terrain affected by selective logging is important to the establishment of credible measures to improve the ecological sustainability of forestry practices. Recent appreciation of the impact of parameter uncertainty on the statistical credibility of complex models with little internal state validation supports the use of more parsimonious approaches such as data-based mechanistic (DBM) modelling. The DBM approach combines physically based understanding with model structure identification based on transfer functions and objective statistical inference. Within this study, these approaches have been newly applied to rainfall-SSflux response. The dynamics of the sediment system, together with the rainfall-river flow system, were monitored at five nested contributory areas within a 44 ha headwater region in Malaysian Borneo. The data series analysed covered a whole year at a 5 min resolution, and were collected during a period some five to six years after selective timber harvesting had ceased. Physically based and statistical interpretation of these data was possible given the wealth of contemporary and past hydrogeomorphic data collected within the same region. The results indicated that parsimonious, three-parameter models of rainfall-river flow and rainfall-SSflux for the whole catchment describe 80 and 90% of the variance, respectively, and that parameter changes between scales could be explained in physically meaningful terms. Indeed, the modelling indicated some new conceptual descriptions of the river flow and sediment-generation systems. An extreme rainstorm having a 10-20 year return period was present within the data series and was shown to generate new mass movements along the forestry roads that had a differential impact on the monitored contributory areas. Critically, this spatially discrete behaviour was captured by the modelling and may indicate the potential use of DBM approaches for (i) predicting the differential effect of alternative forestry practices, (ii) estimating uncertainty in the behaviour of ungauged areas and (iii) forecasting river flow and SSflux in terrain with temporal changes in rainfall regime and forestry impacts.

Forecasting river flow rate during low‐flow periods using neural networks

The pollution in the river Arno downstream of the city of Florence is a severe environmental problem during low‐flow periods when the river flow rate is insufficient to support the natural waste assimilation mechanisms which include degradation, transport, and mixing. Forecasting the river flow rate during these low‐flow periods is crucial for water quality management. In this paper a neural network model is presented for forecasting river flow for up to 6 days. The model uses basin‐averaged rainfall measurements, water level, and hydropower production data. It is necessary to use hydropower production data since during low‐flow periods the water discharged into the river from reservoirs can be a major fraction of total flow rate. Model predictions were found to be accurate with root‐mean‐square error on the predicted river flow rate less then 8% over the entire time horizon of prediction. This model will be useful for managing the water quality in the river when employed with river quality models.

A non-linear perturbation model considering catchment wetness and its application in river flow forecasting

A non-linear perturbation model for river flow forecasting is developed, based on consideration of catchment wetness using an antecedent precipitation index (API). Catchment seasonality, of the form accounted for in the linear perturbation model (the LPM), and non-linear behaviour both in the runoff generation mechanism and in the flow routing processes are represented by a constrained non-linear model, the NLPM-API. A total of ten catchments, across a range of climatic conditions and catchment area magnitudes, located in China and in other countries, were selected for testing daily rainfall-runoff forecasting with this model. It was found that the NLPM-API model was significantly more efficient than the original linear perturbation model (the LPM). However, restriction of explicit non-linearity to the runoff generation process, in the simpler LMP-API form of the model, did not produce a significantly lower value of the efficiency in flood forecasting, in terms of the model efficiency index R 2.

A nearest neighbour linear perturbation model for river flow forecasting

The non-parametric nearest neighbour method (NNM) for river flow forecasting is explored and developed further as the nearest neighbour linear perturbation model (NNLPM), which combines the nearest neighbour concept with the concept of perturbations from a mean value, as used in the linear perturbation model (LPM). The NNLPM model is tested on six catchments and its results are compared with those of the simple linear model (SLM) and the linear perturbation model (LPM). The results indicate that the NNLPM is a more reliable indicator of the discharge forecast than either the SLM or the LPM in the case of non-seasonal catchments.

Parsimony, Model Adequacy and Periodic Correlation in Time Series Forecasting

The merits of the modelling philosophy of Box \& Jenkins (1970) are illustrated with a summary of our recent work on seasonal river flow forecasting. Specifically, this work demonstrates that the principle of parsimony, which has been questioned by several authors recently, is helpful in selecting the best model for forecasting seasonal river flow. Our work also demonstrates the importance of model adequacy. An adequate model for seasonal river flow must incorporate seasonal periodic correlation. The usual autoregressive-moving average (ARMA) and seasonal ARMA models are not adequate in this respect for seasonal river flow time series. A new diagnostic check, for detecting periodic correlation in fitted ARMA models is developed in this paper. This diagnostic check is recommended for routine use when fitting seasonal ARMA models. It is shown that this diagnostic check indicates that many seasonal economic time series also exhibit periodic correlation. Since the standard forecasting methods are inadequate on this account, it can be concluded that in many cases, the forecasts produced are sub-optimal. Finally, a limitation of the arbitrary combination of forecasts is also illustrated. Combining forecasts from an adequate parsimonious model with an inadequate model did not improve the forecasts whereas combining the two forecasts of two inadequate models did yield an improvement in forecasting performance. These findings also support the model building philosophy of Box \& Jenkins. The non-intuitive findings of Newbold \& Granger (1974) and Winkler \& Makridakis (1983) that the apparent arbitrary combination of forecasts from similar models will lead to forecasting performance is not supported by our case study with river flow forecasting.

Forecasting river flow using nonlinear dynamics

A Nearest Neighbor Method (NNM) is used to forecast daily river flows that were measured at a single location over a time period spanning about seventy years. A parsimonious three parameter NNM is developed in the context of Nonlinear Dynamics and the dependence between forecast error and length of history used to construct forecasts is investigated. Comparison is made to Auto-Regressive Integrated Moving Average (ARIMA) models. The NNM is found to provide improved forecasts.