Design of an adaptive neuro-fuzzy computing technique for predicting flow variables in a 90° sharp bend

The presence of bends causes complex flow behaviour in rivers and open channels. Secondary and spiral flows are the main features in curved channels (Lien et al. 1999; Naji et al. 2010). The complex and three-dimensional nature of flow leads to change the velocity and water surface profile distribution (Ferguson et al. 2003; Zeng et al. 2008). Shukry (1950), Rozovskii (1961) and Anwar (1986) carried out the first extensive experimental studies on 90° and 180° sharp and mild bends. Rozovskii (1961) studied the flow properties and distribution of boundary shear stresses in curved channels with 90° mild bends and 180° sharp bends. Rozovskii also examined the location where the maximum velocity occurs and its displacement. He found that with mild bends, the maximum velocity displacement begins in the internal cross sections of the channel and is transmitted to outer walls because secondary flows overcome the pressure gradient. The velocity distribution and causes of secondary and spiral flows forming in bends were experimentally evaluated by Ye & McCorquodale (1998), Blanckaert & Graf (2001), Barbhuiya & Talukdar (2010) and Uddin & Rahman (2012). DeVriend & Geoldof (1983), Steffler et al. (1985), Bergs (1990) and Gholami et al. (2014) experimentally investigated transverse and longitudinal water surface profiles and slopes. In the numerical field, Leschziner & Rodi (1979), Bodnar & Prihoda (2006) and Ramamurthy et al. (2013) investigated the flow patterns in a 90° sharp bend and reported a nonlinear water surface slope. DeMarchis & Napoli (2006) numerically studied the distribution pattern of velocity components in a 270° bend. Their results indicated that the end section of the mild bend causes greater velocity at the outer channel wall than in other sections. Bonakdari et al. (2007) worked with a computational fluid dynamics (CFD) model to survey the bending effect on the velocity pattern in a channel with circular cross sections. Vaghefi et al. (2015) used the depth-averaged method to study and analyse the shear stress distribution near the bed in a 180° sharp bend. Gholami et al. (2014) experimentally and numerically investigated the flow pattern in a 90° sharp bend with focus on the general flow characteristics. They evaluated the models’ performance and found that sharp bends exhibit a full, three-dimensional, complex flow pattern.

In recent decades, artificial intelligence (AI) methods have become popular flow forecasting means in many engineering sciences. Numerous researchers have applied AI to solve an array of hydraulics problems, such as stream flow forecasting (Chiang et al. 2004; Chau & Wu 2010) and flow in open channels (Yuhang & Wenxin 2009; Gholami et al. 2015, 2016a, 2016b), and scour and sediment transport prediction (Liriano & Day 2001; Guven & Gunal 2008; Ebtehaj & Bonakdari 2013). AI methods are also frequently employed to predict the flow variables in curved channels. Sahu et al. (2011) used artificial neural network (ANN) modeling to study and predict the velocity values within a meander open channel. Bonakdari et al. (2011) predicted the velocity field values in a 90° mild bend using the multilayer perceptron neural network and genetic algorithm (GA). Their results indicated that the proposed networks are able to predict velocity values at points where experimental data are not available. Baghalian et al. (2012) examined ANN model performance and compared it with an analytical solution. They also studied a CFD model of the flow patterns in curved channels. According to their results, the multi-layer perceptron (MLP) and CFD models predicted the flow pattern well.

More recently, fuzzy logic and neural networks have led to a novel neuro-fuzzy system called the adaptive neural fuzzy inference system (ANFIS). Asadiani Yekta et al. (2010) examined suspended sediment estimation using ANFIS, ANN and sediment rating curves. Their results showed that ANFIS is more accurate than ANN in estimating suspended sediment load as a discharge function. Samandar (2011) presented an ANFIS model to predict the friction coefficient of open channel flow. ANFIS modeling may be a suitable method of analysing general hydraulics problems that are mostly based on laboratory tests. Bahramifar et al. (2013) predicted the Manning roughness coefficient in alluvial channels using ANFIS. Moharana & Khatua (2014) applied ANFIS modeling to predict the roughness coefficient in meander open channel flow. The results signified that ANFIS modeling is a suitable and effective method of predicting the nonlinear relationship between the roughness coefficient and effective non-dimensional factors. Ebtehaj & Bonakdari (2014) used ANFIS models with different training algorithms to predict sediment transport in sewers. By comparing the ANFIS models with theoretical equations, they concluded that ANFIS models can make more accurate predictions.

No study has been done on ANFIS in terms of evaluating the flow parameters in channel bends. Therefore, in this study, the velocity and water surface depth parameters of a 90° sharp bend were applied using the ANFIS method. In order to achieve FIS generation, sub-clustering (SC) and grid partitioning (GP) were employed and for model training, the back propagation (BP) algorithm and a hybrid of BP and least squares (LS) were used. As such, four different states (ANFIS-GP-BP, ANFIS-GP-Hybrid, ANFIS-SC-BP and ANFIS-SC-Hybrid) were applied in the current ANFIS design. Various network models were trained and tested based on extensive experiments in a 90° sharp flume. The velocity and water depth data were all related to five different discharge values. The inputs to each model were the coordinates (X and Y) of different points in the bend and discharge (Q) rate, and the outputs were velocity and flow depth, respectively.

To examine the FIS generation in this study, the GP and SC methods are used. The performance of each method is expressed next.

In this approach, the input space is apportioned in several rectangular subspaces using a number of fuzzy regions. The space partitioning is achieved based on the type and number of MFs in each dimension by the portion parallel to the axis. To determine the fuzzy sets and parameters, the LS approach according to the membership function type and partition is utilized. It should be noted that the consequent parameters related to the linear output of the membership function are set to zero during the production of fuzzy rules. Subsequently, the parameters are determined and refined by ANFIS. The ANFIS and GP combination was presented by Kennedy et al. (2003) in detail. The number of fuzzy rules is exponentially dependent on the number of input parameters. By considering that the number of input parameters and the membership function for each parameter are m and n, the whole number of fuzzy rules is equal to mⁿ. Therefore, GP is not suitable for problems with large numbers of input variables (Jang 1993). The number of input variables in this study is equal to 3 (X, Y, Q), which allows using GP in FIS generation.

The SC approach is an extension of mountain clustering (Yager & Filev 1994). In this method, data points are utilized in GP instead of grid points, and each data point can be considered a potential cluster center. The effective grid point that must be evaluated is independent of the problem dimension and is equal to the number of data points. Furthermore, this technique does not require specifying the grid resolution, where tradeoffs between accuracy and computational complexity would have to be taken into account. The SC process is presented as follows.

The procedure of obtaining a new cluster center and revising the potential of each is repeated until the residual potential of all data points is lower than some deduction of the first cluster center's potential ⁠.

The influential radius value is directly dependent on the amount of rules, such that a small radius results in small clusters and a higher number of rules and vice versa. Consequently, the influential radius needs to be properly selected for data space clustering. The algorithms applied for training include BP and a hybrid of BP and LS. Thus, the ANFIS network design has four different states. The network inputs include X, Y and Q (longitudinal and transverse coordinates and discharge, respectively). A single output is considered, either the velocity or flow depth parameter. As a result, eight different models are presented in this study.

Bowden et al. (2002) optimized the data used for ANN model training and testing by two methods: the GA and self-organizing map (SOM). GA was applied to a problem of dividing the data into three statistically similar subsets. The SOM technique achieves an equally distributed training set by first dividing all data into a number of classes. Then by sampling an equal number of records from each cluster, the training, testing and validation set data are distributed throughout the problem domain equally. Bowden et al. (2002) stated that the poor performance of the ANN model was related to their own data and not the choice of ANN model parameters or architecture. The two mentioned methods reduced the prediction error pertaining to the conventional data division method. The method for data selection was such that from 520 velocity data, 400 were used for model training (75%) and the remaining 25% (120 data) were used for model performance testing. From 506 different flow depth data, 386 were employed for model training and 120 for model testing.

With the GP-Hybrid model, data compression around the fitted line was greater than the other models, and it can thus be concluded that GP-Hybrid predicted velocity more accurately. It is not obvious which model was more accurate from these figures. Therefore, Table 3 shows the statistical indices for the GP-Hybrid, GP-BP, SC-Hybrid and SC-BP models with the training and testing datasets. The training stage was more accurate than the testing stage. However, there is no significant difference between the results from the two stages, and in the worst condition the SC-Hybrid model's accuracy (R² value) decreased by 16% in the testing stage. The table demonstrates that the GP-hybrid model has the highest R² error index value compared with the other models. The high R² value indicates the good agreement between the values predicted by this model and experimental data. The lower RMSE and SI index values for the GP-BP model are obvious, but the nearly equal values for these indices to GP-Hybrid represent the similar nature of these two models.

However, GP-Hybrid is introduced as the superior model owing to the higher R² and MARE (0.867 and 0.076, respectively) than the GP-BP model. The relative error index of MARE increased by nearly 6% by changing the FIS generation method to SC, meaning that the SC-Hybrid model was less accurate than the GP-Hybrid model. In the GP-Hybrid model, however, by changing both the FIS generation method and training algorithm, the MARE error index increased by 12%. Hence, the SC-BP model was less accurate in predicting the velocity parameter due to the high SI error index (SI = 0.099). Therefore, it can be said that the GP generation method is more accurate than the SC method. Moreover, for both generation methods (GP and SC), using hybrid models is better than the BP training algorithm. The positive and negative bias index values also indicate the overestimation and underestimation of these models, respectively. According to the negative and positive values in Table 3, it can be said that the GP-BP and GP-Hybrid models underestimated and overestimated, respectively.

As a result, the GP-Hybrid model is considered the most accurate in velocity prediction. In the grid-partitioning model for FIS generation with the training hybrid algorithm, the R² value in both stages (training = 0.968, testing = 0.867) was the highest. In training, the GP-Hybrid model had the lowest error values (RMSE = 1.378, MARE = 0.027) compared with the other models. The bias and SI values for this model were 0.079 and 0.09, respectively.

Four models, i.e. GP-Hybrid, GP-BP, SC-Hybrid and SC-BP were designed separately to predict water depth. The evaluation of the four models using different statistical indices for the training and testing datasets is shown in Table 4. Evidently, all models had the same R² value and the same function; hence, in this study, different discharge rates (different water depths) were considered as input parameters and the results of each ANFIS model are given in terms of total discharge.

Similar to the velocity prediction models, GP-Hybrid had lower error values in the testing stage (RMSE = 0.155, MARE = 0.02) and outperformed the other models. Moreover, the bias and SI indices of 0.107 and 0.017 respectively, of this model were lower than the comparable values of the other model. The GP-Hybrid model predicted water depth in a 90° bend reasonably well.

By changing the training algorithm from hybrid to BP, the model error increased and the GP-BP model had the greatest relative and absolute errors (RMSE, MARE and SI of 0.022, 0.028 and 0.024, respectively) and was thus introduced as the least accurate model for water surface depth prediction. By changing the generation method from GP to SC, the SC-BP and SC-Hybrid models’ relative error (MARE) slightly increased by 3.5% and 4.5%, respectively, but SC-BP (with SI = 0.0170) was more accurate than the SC-Hybrid model (SI = 0.0173). Therefore, it can be said that unlike the velocity prediction models, the SC-BP model with MARE index of 0.16 ranks second after the GP-hybrid model in terms of flow depth prediction accuracy; the SC-BP model performed similar to the GP-Hybrid model with equal MARE and SI error index values. The negative bias index values for the four models indicate underestimation, especially from the two SC models, due to the negative bias index values in the training stage.

In the sections before the bend the flow velocity distribution is uniform, similar to straight channels. As flow enters the bend, the maximum velocity occurs at the inner channel wall. By advancing into the bend and with the power of secondary flows, the maximum velocity gradually transmits to the channel axis and transfers to the outer wall in the cross sections located after the bend. In the sections after the bend, the flow velocity is not yet uniform and the maximum velocity is at the outer channel wall. In the sections after the bend, the velocity profiles will be uniform after flow establishment length. However, during the experiments, the mean velocity profiles 20–30 cm before and 150–160 cm after the bend remained constant and a fully developed zone (establishment length) was achieved in our experiments. Rozovskii (1961) stated that the main factor behind velocity displacement in mild bends is secondary flows, while in sharp bends it is the longitudinal pressure gradient; therefore the establishment length in mild bends is more important than in sharp bends, due to the power of secondary flows. He offered a relation for determining a specific length when the secondary flow has maximum strength. Based on this relation, he concluded that in order for secondary flow to develop, there needs to be a bend with a central angle of at least 100° for shallow flumes, and a 180° central angle for deep flumes. He also observed that the logarithmic distribution probability for velocity profiles leads to favourable results.

In this study, the velocity and flow depth parameters in a 90° sharp bend were predicted using ANFIS models. In the ANFIS design, two methods were used for FIS generation, namely GP and a SC algorithm, while BP and a hybrid of BP and LS were used to train the models. Thus, ANFIS modeling of the velocity and flow depth parameters was done with four different models: GP-BP, GP-Hybrid, SC-BP and SC-Hybrid. In each model, coordinates of 13 points in eight different cross sections and different discharge (Q) rates were considered as input parameters. The modeling results indicate that the velocity and flow depth prediction by the GP method in a 90° sharp bend was more accurate than SC for FIS generation. Additionally, for each velocity and flow depth prediction, the GP model with a hybrid of BP and LS training algorithm (R² of 0.867 and 0.998, respectively) outperformed the backpropagation algorithm. The error contours in the bend also show that all four models achieved the highest accuracy at the inner wall (separation zone) and lowest accuracy at the outer wall (contraction zone). For future work, it is recommended to build an ANFIS model using an evolutionary algorithm such as the GA and differential evolution and compare its performance to the model with BP and hybrid algorithm employed in this study.