SEDIMENT FLOW MODELING IN BHAGIRATHI RIVER USING ARTIFICIAL NEURAL NETWORKS

Abstract

A river system is a complex network of intertwining channels with an ongoing interaction of flow and sediment transport processes. The prediction of sediment transport is of vital interest due to its importance in understanding river hydraulics, geomorphology, irrigation, hydropower, design and management of water resources projects etc. Hydrodynamic models have been widely used for the analysis, prediction, design, and management of a wide range of water– sediment systems. However, due to the spatial heterogeneity of various physical and geomorphologic properties, a river system cannot be easily represented, and data requirements are large for modeling. A river network covers a vast area comprising of many watersheds and sub- basins, where a complete set of data may not be available. This results in the need for a practical userfriendly model which enables quick simulations and predictions with minimum data requirement and without significantly compromising the model accuracy. ANN has the characteristics of parallel link, error correction, and nonlinear transfer and is an emerging technique for the flow and connection of information. The advantage of using ANN is that every step of the modeling process can be configured and improved based upon model performance. This increases flexibility and also the understanding of the procedure which is otherwise rather complex to comprehend. In the present study, Artificial Neural Networks (ANN) have been employed to model sediment flow in the Himalayan Bhagirathi River, using multi annual time series data at four locations viz. Gangotri, Maneri, Uttarkashi and Rishikesh (Ganges River). ANN modeling in the present study has been carried out by- understanding the geohydrological processes and parameters which control the variations in sediment load in the Bhagirathi River, description and analysis of time series data, development of representative models, training, testing and evaluation. The possibility of modeling sediment concentration with Artificial Neural Networks at Gangotri, the source of Bhagirathi River has been explored. Considering discharge, rainfall and temperature to be the main controlling factors of variations in sediment concentration in the dynamic glacial environment of Gangotri, fourteen feed forward neural networks with error back propagation algorithm with different inputs have been created, trained and tested for prediction of sediment concentration. The inputs applied in the models are either the variables mentioned above as individual factors (single input networks) or a combination of them (multi-input networks). The suitability of employing antecedent time-step values as network inputs has been checked by comparative analysis of model performance in two different iii modes. The simple feed forward network has been improvised with a series parallel NARX [nonlinear autoregressive with exogenous input] architecture wherein true values of sediment concentration have been fed as input during training. Daily data of discharge, rainfall, temperature and sediment concentration for the melt period of May-October, when maximum sediment movement takes place, for five years, from the year 2000 to 2004, has been used for modeling and high Coefficient of Determination values [0.77-0.88] have been obtained between observed and ANN predicted values of sediment concentration. According to the performance parameters (R and R2 values), among discharge, temperature and rainfall as independent variables, the sediment concentration is most affected by rainfall (highest R2 value). In this scenario, according to the performance parameters, the rainfall-temperature (T7) combination of inputs is seen to work relatively better. The overall performance range is not too large with the values of coefficient of determination ranging from 0.777 to 0.885. This implies that the generally accepted belief of better performance of multi-input ANN models may not hold true always. It is also seen that use of previous time step values as inputs (updating mode) may not necessarily improve model performance and vice versa. The study has brought out relationships between variables that are not reflected in normal statistical analysis. A strong rainfall: sediment concentration and temperature: sediment concentration relationship is shown by the models which is not reflected in statistical correlation. It has also been observed that usage of antecedent time step values as network inputs does not necessarily lead to improvement in model performance. At Maneri, a simple technique for prediction of suspended sediment concentration [SSC] is presented. ANN models have been developed using short time period data of discharge and sediment concentration during the high activity monsoon period of June to October, 2004, when variations are maximum. Two modeling approaches have been employed, a daily approach and a three hourly approach. Although the time period considered is the same in both the approaches, the modeling performance is marginally better in the three hourly approach where there is a six fold increase in the dataset. The Levenberg-Marquardt optimization function, improvised with NARX [non-linear autoregressive with exogenous input] architecture has been used and high values of coefficient of determination have been obtained [0.89-0.97]. This study shows that short duration time series data can be used for successfully predicting geo-hydrological variables in the highly complex Himalayan river scenario. Single series modeling using Nonlinear Autoregressive networks has been carried out considering six water years discharge data at Maneri, Uttarkashi and Rishikesh. In this form of iv modeling, the present time discharge values were predicted using antecedent discharge values. High values of coefficient of determination were obtained in the study [R2=0.92-0.95]. The study validates the possibility of single series prediction in the Bhagirathi River. Prediction would not just help in filling gaps in hydrological data but will also enable continuous monitoring of sediment concentration which is difficult in the Himalayan Rivers where floods and other such eventualities commonly occur. The present work, besides validating the use of ANN in geo-hydrological modeling, discusses several finer nuances of this technique. For instance, the pros and cons of- using single/multiple inputs in models, the use of previous time step values as network inputs, long duration data vs short duration data, daily vs high frequency three hourly data, the importance of data normalization, pre- modeling data analyses with statistical methods, prediction with a single series etc. Such a study would be of great help in understanding the relationships that exist between hydrologic variables and the degree to which they affect sediment movement. ANN prediction can also be useful for filling up the gaps in hydrologic data. This study shows that geohydrological time series data invariably has inherent trends which can be exploited for ANN modeling and predictions.