**Effect of Sediment Load Boundary Conditions in Predicting Sediment Delta of Tarbela Reservoir in Pakistan**

**Zeeshan Riaz Tarar 1, Sajid Rashid Ahmad 1, Iftikhar Ahmad 1, Shabeh ul Hasson 2,3, Zahid Mahmood Khan 4, Rana Muhammad Ali Washakh 5,6, Sardar Ateeq-Ur-Rehman 4,7,\* and Minh Duc Bui <sup>7</sup>**


Received: 16 June 2019; Accepted: 13 August 2019; Published: 18 August 2019

**Abstract:** Setting precise sediment load boundary conditions plays a central role in robust modeling of sedimentation in reservoirs. In the presented study, we modeled sediment transport in Tarbela Reservoir using sediment rating curves (SRC) and wavelet artificial neural networks (WA-ANNs) for setting sediment load boundary conditions in the HEC-RAS 1D numerical model. The reconstruction performance of SRC for finding the missing sediment sampling data was at R2 = 0.655 and NSE = 0.635. The same performance using WA-ANNs was at R<sup>2</sup> = 0.771 and NSE = 0.771. As the WA-ANNs have better ability to model non-linear sediment transport behavior in the Upper Indus River, the reconstructed missing suspended sediment load data were more accurate. Therefore, using more accurately-reconstructed sediment load boundary conditions in HEC-RAS, the model was better morphodynamically calibrated with R<sup>2</sup> = 0.980 and NSE = 0.979. Using SRC-based sediment load boundary conditions, the HEC-RAS model was calibrated with R2 = 0.959 and NSE = 0.943. Both models validated the delta movement in the Tarbela Reservoir with R2 = 0.968, NSE = 0.959 and R2 = 0.950, NSE = 0.893 using WA-ANN and SRC estimates, respectively. Unlike SRC, WA-ANN-based boundary conditions provided stable simulations in HEC-RAS. In addition, WA-ANN-predicted sediment load also suggested a decrease in supply of sediment significantly to the Tarbela Reservoir in the future due to intra-annual shifting of flows from summer to pre- and post-winter. Therefore, our future predictions also suggested the stability of the sediment delta. As the WA-ANN-based sediment load boundary conditions precisely represented the physics of sediment transport, the modeling concept could very likely be used to study bed level changes in reservoirs/rivers elsewhere in the world.

**Keywords:** Upper Indus Basin (UIB); Tarbela Reservoir; Besham Qila; sediment modeling; uncertainty; wavelet transform analysis-artificial neural network (WA-ANN); sediment rating curve (SRC); HEC-RAS

#### **1. Introduction**

The uncertainties in modeling reservoir sedimentation are due to: (a) both flow and sediment; (b) the distribution of sediment particle size; (c) the specific weights of sediment deposits; (d) reservoir geometry; and (e) the operational rules of reservoirs [1]. These uncertainties are propagated, particularly, due to the varying input of sediment loads as boundary conditions. Normally, sediment series, as input to the model, are estimated by utilizing sediment rating curves (SRCs), prepared by developing relationships through simple regression techniques, between flow and sediment, observed over a considerable period, adequately representing the complete hydrological cycle over decades [2,3]. It has been observed in various sediment studies of reservoirs around the world that SRCs, though a simple and convenient way to estimate missing values of sediment inflow, often overestimate and overshoot the sediment entry into the reservoirs against the actual conditions, up to 50% [4,5]. Tarbela Reservoir hydrographical/bathymetric surveys have been conducted since 1979 to observe the sediment entry and position/advancement of the delta in the reservoir. Each year, the reservoir authorities issue Sedimentation Reports based on the above conducted surveys. As per the Sedimentation Report of Tarbela Reservoir [6], the actual observed sediment deposits in the reservoir are about 171.3 Mt/year, which are about 53% of the average of the below-mentioned studies, i.e., 47% overestimation. Hence, precise hydro-morphodynamic boundary conditions play a principal role in modeling the transport processes in rivers and reservoirs.

The Tarbela Dam Project (TDP) was completed in the mid-1970s and is the backbone of the hydropower and water resources of Pakistan, with its 3478 MW of existing installed and 6298 MW of near future capacity. It is the world's largest earth-filled dam and also by structural volume [7]. The Tarbela Reservoir drains UIB and lies at its lowest point. The drainage area up to the dam is about 170,000 km2, as shown and demarcated in Figure 1. The huge body of water created behind the dam, originally 11.620 million acre feet (MAF), has been reduced by sedimentation to 6.856 MAF in 2019 [8], meaning that it is only 59% of its original storage volume, and the rest has been consumed by sedimentation. The feasibility and engineering studies of Tarbela Dam that were conducted in the mid-1960s and 1970s took serious note of the potential sedimentation problems that were likely to arise after some years of dam construction. Various studies at the time and afterwards estimated sediment entering the reservoir to be substantially overestimated, based primarily on techniques in vogue and with less data. The Tarbela Dam Consultants (Tippets, Abbett, McCarthy, Stratton (TAMS)) used 235 million tons (Mt) annually as the sediments entering the reservoir [9]. The Kalabagh Dam Consultants estimated the annual sediment load entering Tarbela as 295.7 Mt using sediment rating curves. The same figure of 295.7 Mt was adopted for sediment studies of Tarbela by the Consultants of the Ghazi Barotha Hydropower Project located just 8 km downstream of Tarbela Dam. The Consultants for the Mega Diamer-Basha Dam, making use of additional data from 1962–2003 in sediment rating curves, calculated the load for Tarbela Reservoir as 233 Mt annually [10]. Future sedimentation scenarios fir Tarbela Reservoir hold a pivotal position for authorities and water managers alike, as a reduction in the storage capacity of Pakistan's largest water body and its implications for all related disciplines would be sensitive enough to provoke studies into alternative or preventive measures.

A list of studies also cited by [11], in addition to the ones mentioned above, calculating sediment entering Tarbela Reservoir/main Upper Indus Basin (UIB), is tabulated in Table 1:


**Table 1.** Published estimates of sediment load (SL) of the Upper Indus River.

**Figure 1.** Demarcated Upper Indus Basin at Tarbela Dam.

All above estimates were based on sediment rating curve (SRC) method and varied in a wide range from around 200 Mt y−<sup>1</sup> – 675 Mt y−<sup>1</sup> over the last 50 years. Unfortunately, the accuracy of SRCs is limited, as they map all scattered data points of discharge and sediment loads using a single fitting line, which is more likely to be affected by data outliers [22–24]. Therefore, the single fitting line cannot handle sediment transport processes connected to the phenomenon of hysteresis and noticeable hydrological variations, such as: (a) fluvial erosion and transport processes, interacting with other sediment-production processes; (b) sediment temporary storage in the main channel of the river [25]; (c) landslide phases related to aggradation and degradation [26]; (d) on average, 5–10 waves of high flow of an average of 10–12 days' duration during the monsoon period; (e) different discharge and sediment conveyance times and their differing lag-times from sources to the gauge recording stations. Basically, all these processes cause different sediment concentrations on same magnitude of discharge during rising and falling limbs of flood events, which is referred to as the hysteresis phenomenon. As SRCs are mostly employed in the estimation process of sediment load boundary conditions due to their construction simplicity, a marked compromise could arise in the numerical or physical modeling outcomes.

Since the variations in sediment load boundary conditions affect the calculations of the morphodynamics, it is essential to model time-related changes in sediment supply more accurately, influenced by the above-mentioned phenomenon of hysteresis and noticeable hydrological changes. During recent years, artificial neural networks (ANNs) have gained increased reception as new analytical techniques due to their robustness and ability to model non-stationary data series. Therefore, ANNs have a clear advantage over other conceptual models as they do not need previous knowledge of the process because they build a relationship between data inputs and targets using non-linear activation functions. The ANNs have multiple inputs with dissimilar characteristics, making ANNs be able to represent time-space variation [1]. In spite of the adequate flexibility of ANNs in modeling time series, sometimes, ANNs have a weakness when signal alterations are highly non-stationary and physical hydrological processes operate under scales of large ranges, with variations of one day to several years. In such a situation, different methods have been proposed, among which are wavelet transforms. They have become a capable method for analyzing such changes and trends in hydrological time series [27–31]. A wavelet has been defined as a small wave whose energy is limited in a short period of time and is a logical method for signals that are non-stationary, having short-lived transient components, featuring at different scales, or singularities. A non-stationary signal can be broken up into a certain number of unvarying signals by wavelet transform. ANN is then combined with wavelet transform (WA-ANN). It is considered that WA-ANN models are more precise than the conventional methods since wavelet transforms provide effective break-ups of the original time series, and the wavelet transformation data improves the performance of conventional ANN models by catching effective information for various resolution levels [4,5,11].

In the present study, effort has been made to model the sediment delta of Tarbela Reservoir using the 1D HEC-RAS numerical model with the objective to reduce variations in its future prediction by employing first the conventionally-estimated sediment inflow based on SRC and then by the above elaborated innovative WA-ANN technique. The sediment series based on WA-ANN, as developed by [4,11], was further updated, calibrated, and validated by inclusion of sediment data up to 2014 and used as input to the model.
