Daily suspended sediment forecast by an integrated dynamic neural network
Introduction
The presence of suspended sediment in river flows is an essential issue in terms of water quality, river geological and geographical settings, channel navigability, operation of hydraulic structures, river aesthetics and fish habitats (Kişi, 2006). For instance, from the perspective of water resources exploitation, suspended sediment with high turbidity in river flows is a physical pollutant, which may act as a chemical pollutant when carrying chemicals such as phosphorous and heavy metals (Doğan et al., 2007). The transport of suspended sediment also affects erosion and deposition, leading to river morphological changes (Liu et al., 2013).
Defined as the ratio of sediment mass to volume of a sediment-laden mixture, suspended sediment concentration (SSC) characterizes a sediment transport process (Zheng et al., 2011). Accurate predictions of the SSC are of practical implications in a wide range of hydraulic issues, including design of the reservoir dead storage, environmental impact assessment, and prediction of aggradation and degradation around bridge piers (Afan et al., 2016). However, the transport is a complex phenomenon, involving both the flow and sediment motions and their interactions. From the initial detachment of sediment particles to their deposition, every stage is characterized by high nonlinearity and interplay processes, rendering the SSC forecast a challenging task.
Conventionally, the SSC in a stream is determined via either direct field measurements or already established sediment transport equations (Bayram et al., 2012). The former is an accurate approach, with however some practical and cost disadvantages. The latter refers to the use of the relationship between the SSC and flow properties, including flow velocity, discharge, etc. The approaches are usually developed based on simplified partial differential equations, assumptions and empirical correlations for erosive effects of rainfall and flow (Aytek and Kişi, 2008, Karim and Kennedy, 1990, McBean and Al‐Nassri, 1988).
Some sediment curves are also proposed for the estimation of SSC (Crowder et al., 2007, Petersen-Øverleir, 2004), which are theoretically adequate to cope with the variations in catchment properties and uneven distribution of precipitation and evapotranspiration (Aytek and Kişi, 2008). However, due to the scarcity of detailed spatial and temporal hydrological data, applications of those methods suffer from limitations (Kisi et al., 2012). Moreover, despite several established empirical models, choosing an appropriate one for a specific scenario is difficult, which is the reason why many of them do not gain expected acceptance (Afan et al., 2016). Some researchers even indicate that the sediment curves are misleading and the high goodness of fit is spurious (McBean and Al‐Nassri, 1988). As a result, efforts have been devoted to develop reliable methods for SSC predictions. Machine learning modeling offers an alternative.
In comparison with process-driven methods, machine learning models gain superiority in dealing with nonlinear issues and handling a large quantity of data. With the help of artificial intelligence (AI), they demonstrate a great potential in mapping data with high complexity, dynamism and non-stationarity, e.g. SSC time series. The recent years have witnessed a rapid growth of AI-based approaches in studying the SSC. One of the first applications is made by Jain (2001), where an integrated relation among water stage, flow discharge and SSC is established using a feedforward neural network (FFNN). The resulting model leads to better predictions than the conventional curve-fitting approach and is capable of modeling the hysteresis effects caused by unsteady flows in channels. Alp and Cigizoglu (2007) examine the performance of the FFNN and radial basis function neural network (RBFNN) in sediment load estimations. The combination of rainfall and flow data generates, as predictors, accurate results; the two evaluated models exhibit insignificant differences. Melesse et al. (2011) adopt a multilayer perceptron neural network (MLPNN) and use precipitation, river flow, and antecedent sediment data to predict the suspended sediment in three major U.S. rivers (Mississippi, Missouri and Rio Grande). Compared with the multiple linear and nonlinear regressions and autoregressive integrated moving average, the MLPNN produces more satisfactory results. Demirci and Baltaci (2013) explore the efficiency of fuzzy logic models in SSC estimations, concluding that they behave better than the multiple linear regression method. Ehteram et al. (2021) use the whale algorithm, particle swarm optimization and bat algorithm to optimize the performance of an artificial neural network (ANN). The hybrid models cut down the error up to 80% relative to the standard one. Doroudi et al. (2021) integrate the observer-teacher-learner-based optimization method into a support vector machine, which outperforms all the evaluated ones. Afan et al. (2016) provide a partial review of the relevant studies in the area.
Due to its stochastic temporal variations, time series modeling is a challenging task. Direct simulations might be inadequate to capture its features in both time and frequency domains. Consequently, for better representation, data pre-processing techniques (e.g. sampling, transformation, de-noising and normalization) are suggested to reformulate and reshape the raw signals. In a rainfall forecast study, Mehr et al. (2019) remove the non-stationary features by performing square root and standardization of the original data. Such a procedure produces a weak stationary signal that is easy to model, as the trend in the mean is separated and the variance is suppressed. For better estimation of water table fluctuations, Jeong and Park (2019) suggest a framework for detrending, deseasonalization, normalization and outlier exclusion. This method, by assuming consistency of the trends and stationarity of seasonal variability, detects, in an effective way, the process changes that occur in a groundwater system. If coupled with machine learning models, it improves the efficiency of the standalone ones. Liu et al. (2013) apply a wavelet for signal decomposition and integrate it with an ANN for SSC forecast, showing that wavelet helps better capture highly nonlinear and non-stationary SSC. A similar study by Rajaee et al. (2011) illustrates that the model is capable of obtaining useful information on various resolution levels. Shiri and Kişi (2012) combine wavelet with gene expression programming (GEP), neuro-fuzzy (NF), and ANN for sediment load prediction. It is concluded that the wavelet-based models significantly increase the accuracy of single ones.
These papers demonstrate the importance of data pre-processing for time series modeling. In most works, the main regressors are static models, e.g. ANN and NF. They are suitable for fitting problems, and a more robust model is required for time series forecast (Nanda et al., 2016). For a better SSC estimation, this paper establishes a nonlinear auto-regressive with external inputs (NARX) network and couples it with a new wavelet-based data pre-processing module. The NARX, a dynamic recurrent neural network, is a superior network for handling non-stationary issues. The data pre-processing framework takes the advantage of the discrete wavelet transformation for time and frequency analysis and the multigene genetic programming (MGGP) for feature selection, expecting to generate informative and noise-free inputs for the regressor. Time series data of daily discharge, water level and SSC during 2004–2012 collected at the Cuntan station on the upper reaches of the Yangtze River are used to test the proposed integrated NARX (INARX). Its performance is assessed statistically and compared with that of empirical and conventional data-driven models. The objective of this study is to put forward an improved neural network for real-time and multistep forecasts of the river suspended sediment. A robust prediction framework has profound engineering implications.
Section snippets
Methodology
Two static models, MLPNN and MGGP, are first introduced. As an essential element in the INARX, wavelet transformation (WT) is described. The INARX is established by integrating the WT and MGGP to the NARX, allowing improved multistep forecasts. The evaluation criteria are finally stated.
Study site
To evaluate the developed INARX model, the Cuntan hydrological station on the upper reaches of the Yangtze River is chosen as the study site. The river basin covers an area of 1.8 × 106 km2, accounting for approximately one-fifth of the land area of China (Fig. 5). Originating in the Qinghai-Tibet Plateau (about 5100 m above the mean sea level), the river flows eastwards through a number of major cities including Chongqing and finally into the East China Sea in Shanghai. Globally, it ranks
Results and discussion
With the time series of the hydrological data, the MLPNN, MGGP and INARX models are assessed and compared. In the first two methods, a short-term memory structure is used to improve their learning performance. With signal decomposition by wavelet functions and details scaling by the MGGP, the proposed INARX model that combines a feedforward neural network with a non-linear auto-regressive model provides improved river sediment forecasts. All the models are coded in MATLAB (Version R2021a) on a
Conclusions
For the improved forecast of suspended sediment in river flows, this study establishes a hybrid dynamic neural network (denoted as INARX). It is a nonlinear autoregressive network with exogenous inputs (NARX) coupled with a data pre-processing module. The time series data are decomposed by wavelet transformation (WT), resulting in two types of subcomponents: approximation and details. The latter is processed by a multigene genetic programing (MGGP) model to remove redundancy and construct
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
As part of research projects at KTH “Two-phase flow modeling: evaluations and simulations for safer spillway discharge” and “Quality and trust of numerical modeling of water-air flows for safe spillway discharge”, this study is funded by the Swedish Hydropower Center (SVC) and supervised by James Yang and Anders Ansell. SVC has been established by the Swedish Energy Agency, Energiforsk and Svenska Kraftnät together with Luleå University of Technology (LTU), Royal Institute of Technology (KTH),
References (64)
- et al.
A wavelet neural network conjunction model for groundwater level forecasting
J. Hydrol.
(2011) - et al.
Past, present and prospect of an Artificial Intelligence (AI) based model for sediment transport prediction
J. Hydrol.
(2016) - et al.
Suspended sediment load simulation by two artificial neural network methods using hydrometeorological data
Environ. Modell. Software
(2007) - et al.
A genetic programming approach to suspended sediment modelling
J. Hydrol.
(2008) - et al.
The accuracy of sediment loads when log-transformation produces nonlinear sediment load–discharge relationships
J. Hydrol.
(2007) - et al.
Groundwater level forecasting using artificial neural networks
J. Hydrol.
(2005) Application of artificial neural networks instead of the orifice plate discharge coefficient
Flow Meas. Instrum.
(2020)- et al.
Rainfall, runoff, and suspended sediment delivery relationships in a small agricultural watershed of the Three Gorges area
China. Geomorphology
(2011) - et al.
Forecasting river water temperature time series using a wavelet–neural network hybrid modelling approach
J. Hydrol.
(2019) - et al.
Monthly streamflow forecasting using continuous wavelet and multi-gene genetic programming combination
J. Hydrol.
(2018)