**3. Results and Discussion**

In the numerical model, daily reservoir water levels (RWLs) of the Tarbela Reservoir were applied for the downstream boundary condition. At the upstream boundary, we specified daily inflows with corresponding sediment load. Modeled results were compared to observations and evaluated based on the statistical performance parameters like the coefficient of determination (R2), the observations standard deviations ratio (RSR), and the Nash–Sutcliffe Efficiency (NSE).

Actual daily inflows of 14 years (2005–2018) were given as the upper flow boundary condition for running the model with the SRC-based sediment loads and were repeated thereafter up to 2030. For running the model with the WA-ANN-based sediment loads, actual daily inflows from 2005–2018 and thereafter futuristic flows from 2019–2030 as projected by [35] under plausible near-future climatic conditions were applied as upper boundary conditions. Actual daily RWLs of the Tarbela Reservoir were given as the downstream boundary condition up to 2018 and repeated thereafter for both SRC and WA-ANN runs of the model.

To check the performance of the SRC method (Equations (1) and (2)), sediment loads were generated and matched against observed sediment loads. The sediment equations output sediment load in t/day by the input of flow in m3/s. The generated/estimated sediment load was matched against observed sediment load entering the reservoir for that particular day. The observed sediment load was calculated by converting observed sedimentation concentration in mg/L for that day into t/day by carrying out a dimensional analysis. The calculated values of NSE, R2, and RSR were 0.635, 0.655, and 0.076, respectively, amply proving that the SRC technique, although in vogue, predicted output with an unacceptable level of certainty.

Applying the concept of data preprocessing on Besham Qila gauge station's data developed by [41], where he found the best relationship by selecting 70% of the input data for training, 15% for testing, and 15% for validation, we also obtained better results for the time period 1969–2014. The 70%, 15%, and 15% data from the entire available series was randomly selected for training, testing, and validation processes, respectively. It is also worthwhile to mention here that data pre-processing plays an important role where a short duration data series is available; however, our data series of more than 40 years also provided us the best results on even specifying 60% of data for training, 20% for testing, and the remaining 20% for validation. The coefficient of determination (R2) for the training and testing datasets was 0.780 and 0.743, respectively. The Nash–Sutcliffe efficiency (NSE) was also 0.780 and 0.742 for training and testing, respectively. As our ANN trained best using single decomposition on Q(t), the inputs were only detailed and approximated coefficients of discharge without lag-time. The best trained WA-ANN used "tainsig" transfer functions in both the hidden and output layers. The number of hidden neurons in the single hidden layer of ANN was only five compared to seven for the same gauge station in [41]. As the Levenberg–Marquardt algorithm has fast convergence and also performed well for the Indus River [4], it also performed best in our training. The simulations stopped when the difference between the last and second to last simulation was less than 1/1000 or it reached maximum epochs of 1000 iterations. The work in [41] used the data series from 1969–2008 and reconstructed missing data for the Tarbela Reservoir with R2 = 0.773 and 0.794 for testing and training, respectively. The statistical performance of our WA-ANN with a larger data series up to 2014 was slightly better for training data; however, it was slightly lower for testing data, which may be due to the inclusion of the exceptionally high flood of 2010. Similarly, increasing the decomposition levels slightly affected the model performance, which, interestingly, was significantly improved in [41]. In addition, the WA-ANN-generated sediment series showed an annual 160 Mt of suspended sediment load (excluding 10% bed load) entering the Tarbela Reservoir, which was similar to the estimate of [41].
