Application of nature-inspired optimization algorithms to ANFIS model to predict wave-induced scour depth around pipelines

Submarine pipelines are commonly utilized to carry gas and oil in offshore areas and usually lie on erodible seafloors. Wave action may wash out sediment around a pipeline due to a localized increase of bed shear stress, with consequent development of a scour hole that may undermine the stability of the pipeline and eventually lead to its collapse. In fact, the pipeline may become suspended in seawater as the scour develops, making its structure not able to withstand static and dynamic forces. Pipeline failure not only represents an economic loss but may also cause significant environmental consequences. Thus, consideration of the scour phenomenon beneath offshore pipelines is key during pipeline design (Fredsøe et al. 1988; Yasa & Etemad-Shahidi 2014).

Most of the available predictive scour depth formulas in this context are based on laboratory experiments (Lucassen 1984; Sumer & Fredsøe 1990; Çevik & Yüksel 1999). Such regression-based equations are of straightforward use, and they are commonly adopted to estimate scouring depth around pipelines and, generally, any river or marine structures; however, possible scale effects may lead to considerable inaccuracy in predicting scour for large-scale structures in the field (Tafarojnoruz 2012; Tafarojnoruz & Gaudio 2012). To overcome this limitation, numerical models may be developed for local erosion simulation (Zhao & Fernando 2007; Zhao et al. 2018); such studies, however, are still limited in number and are generally dependent on validation against laboratory observations. Furthermore, simulating scour phenomena with a numerical model is computationally burdensome: resolving a three-dimensional scour hole up to the point it reaches equilibrium may take weeks to months of machine time, which is often impractical for any project's purposes.

Soft computing (SC) techniques have been increasingly adopted to analyze and predict various hydraulic phenomena. For instance, the use of group method of data handling (GMDH), applied to the prediction of the longitudinal dispersion coefficient in rivers, offered more accurate estimations compared with the available empirical equations (Najafzadeh & Tafarojnoruz 2016); a study on the calculation of riprap stone size for protection of a steep slope revealed that evolutionary polynomial regression (EPR) is a robust alternative to the empirical mathematical formulations (Najafzadeh et al. 2018); numerous studies demonstrated the capability of artificial intelligence (AI) techniques in predicting local scour depth around hydraulic structures (Najafzadeh et al. 2017; Ebtehaj et al. 2018; Najafzadeh & Kargar 2019).

Scour development adjacent to submarine pipelines is generally caused by the shear stress on the seabed associated with waves, currents, or a combination of both. Previous investigations have used AI techniques to predict current-induced and wave-induced scour depth. For current-induced scouring, the earliest studies focused on artificial neural networks (ANNs)-based and Genetic Programming-based scour depth prediction around pipelines crossing rivers, producing an acceptable prediction performance (Azamathulla & Ghani 2010; Azamathulla & Zakaria 2011). Zanganeh et al. (2011) adopted an optimization-based methodology (i.e., PSO algorithm) to mitigate the shortcomings of an adaptive neuro-fuzzy inference system (ANFIS) model for current-induced scour prediction. Yasa & Etemad-Shahidi (2014) derived scour prediction formulations for live-bed and clear-water scour conditions by combining the model tree (MT) and regression model.

Utilizing some of the AI methods may lead to derive new prediction equations. These equations may generally have more complicated mathematical structures than those resulting from the conventional regression-based approaches, but at the same time may offer more accurate predictions. For instance, Najafzadeh & Sarkamaryan (2018) proposed gene-expression programming (GEP), EPR and MT algorithms to extract mathematical formulations for estimating current-induced scour depth below pipelines. Recent studies showed the capabilities of Multivariate Adaptive Regression Splines and Support Vector Machine techniques in predicting scour depth under pipelines in rivers (Haghiabi 2017, 2019; Parsaie et al. 2019).

For wave-induced scour, the prediction performance of an ANN approach, one of the most common AI models, was assessed in comparison with regression-based formulations (Kazeminezhad et al. 2010). Although this study reported accurate predictions using the ANN approach, its application is not easy to carry out by engineers. To overcome this limitation, an MT was later developed to derive more easily usable predictive equations (Etemad-Shahidi et al. 2011). These studies, as well as other applications of different AI approaches, e.g., GMDH (Najafzadeh et al. 2014a, 2014b), demonstrate the feasibility of SC models to estimate the scour depth caused by waves or currents around submarine pipelines. In particular, the combination of two AI techniques has been shown to enhance the prediction performance: for instance, a GMDH network programmed using a GEP technique provided excellent prediction results for scouring under pipelines (Najafzadeh & Saberi-Movahed 2018).

A common AI approach, ANFIS, combines ANN and fuzzy inference system (FIS) and is widely used to estimate scouring depth around hydraulic structures. Because of the flexibility provided by ANNs, ANFIS models can be trained following a complex mathematical mapping between inputs and outputs within a nonlinear framework. Moreover, the ‘IF and THEN’ rules, embedded in FIS, allow for forecasting the behavior of uncertain systems. In recent years, ANFIS models have been applied to estimate the scouring depth at bridge piers and abutments (Akib et al. 2014; Choi et al. 2017; Moradi et al. 2018), culvert outlets (Azamathulla & Ghani 2011) and long contractions along straight canals (Najafzadeh et al. 2016). It can be therefore expected that the ANFIS technique can satisfactorily be used for the prediction of scour around pipelines as well.

Recently, the nature-inspired algorithms, i.e., ant colony optimization (ACO), particle swarm optimization (PSO), differential evolution (DE) and genetic algorithm (GA) have been introduced for optimization purposes in various engineering problems. Specifically to water resources engineering, ACO algorithm has been used to analyze optimal groundwater long-term strategies (Li & Chan Hilton 2005, 2007); PSO algorithm has been adopted to optimize water distribution networks (Surco et al. 2018), rainfall-runoff forecasting models (Motahari & Mazandaranizadeh 2017) and scour depth estimations (Zanganeh et al. 2011; Najafzadeh 2015).

The progression of AI modeling in the field of hydraulic engineering indicates the limitations of the existed AI models and the enthusiasm for solving those limitations. The primary contribution of the present study is to address the internal tuning parameters that are associated with the ANFIS model. This has been scientifically evidenced over the recent literature, and thus the main motivation of the methodological phase is taken place.

The goal of this study is to enhance the capability of the ANFIS technique to estimate scouring depth under submarine pipelines by combining it with the aforementioned nature-inspired optimization algorithms. The prediction accuracy of the proposed methodology is then quantified and compared with the most recent formulations obtained with stochastic approaches for wave-induced pipeline scour depth (Etemad-Shahidi et al. 2011; Sharafati et al. 2018) using indices of prediction performance.

Although for the estimation of maximum scour depth beneath a pipeline placed above the seabed level other researchers (Sumer & Fredsøe 2002; Mousavi et al. 2009) derived predictive formulas neglecting ⁠, a more accurate scour depth estimation is obtained if all the dimensionless parameters in Equation (7) are included. To this end, based on an MT approach, Etemad-Shahidi et al. (2011) derived a set of mathematical formulations to predict the scouring depth by considering all the parameters in Equation (7).

The recent years have seen a noticeable advancement of SC approaches (Sharafati et al. 2019). Such methods are suitable for solving complex problems characterized by a high level of non-linearity and non-stationarity (Yaseen et al. 2015). Hybridized AI-global optimization models have recently gained popularity (Ghorbani et al. 2017). Among them, the performance of nature bio-inspired models has been the best, owing to the AI models-associated hyper-parameters (Maier et al. 2014).

This study hybridized four nature bio-inspired algorithms, namely PSO, ACO, DE and GA, with an ANFIS model for wave-induced maximum scour depth prediction at pipelines. Figure 1 outlines the developed models in the form of flowcharts.

ANFIS models are very well-established AI models based on fuzzy logic (Jang 1996) and their popularity is because they allow input variables (attributes) to execute numerical approximation of the internal mechanism relationships of a physical phenomenon (Yaseen et al. 2017). In essence, ANFIS models boost the learning capability of a classic ANN, which develops rules to map a set of inputs to an output value based on a set of fuzzy rules presented by Zadeh (1965). The fuzzy logic component is exceptionally beneficial as it aids in optimal solution generation (in terms of prediction performance) from imprecise/noisy input attributes.

The fuzzy logic approach is based on allowing each element of a dataset to fall in a particular class (set) partially, and its membership degree is described by a membership function. Achieving an accurate learning process based on knowledge and related experience requires optimal selection of the shape of the membership functions, and fuzzy rules (Kisi & Yaseen 2019).

The optimization of the rules that map the inputs to an output value is typically done with trial and error procedures, and the ANFIS model utilizing this type of optimization is referred to as ‘classic’ ANFIS model. Such trial and error procedures have the disadvantage of often being time-consuming and can lead to overfitting. These issues can be mitigated by using global optimization algorithms. The methodologies for optimization used in this investigation are outlined in the following sub-sections.

The PSO technique is a bio-based optimizer first introduced by Eberhart & Kennedy (1995); it was inspired by the pattern of movement of natural creatures, such as fish, insects and birds. This methodology models each candidate solution to an optimization problem as a particle flowing in the search domain of the optimization problem. The location (position) and speed of every single particle are adjusted following its own experience and the neighboring particles.

The ACO was first presented by Dorigo & Di Caro (1999) as an optimizer which has undergone several modifications to suit multiple engineering applications (Weise 2009; Afshar et al. 2015; Ajay Adithyan et al. 2018). The ACO algorithm is an optimization technique that is most effective in addressing both dynamic and static problems in the field of engineering (Dorigo et al. 1996; Blum 2005; Dorigo & Blum 2005; Dorigo & Socha 2007).

Although colonies of ants are composed of simple individuals, they are considered to have one of the most well-organized structures in nature (Guo & Zhu 2012). The stigmergy mechanism which facilitates self-organization controls activities like foraging, brood sorting, co-operative transport and division of labor (Dorigo & Di Caro 1999). Stigmergy in ant colonies is based on the pheromone track left by each ant, which affects the actions of all the other ants. The ACO algorithm, inspired by this concept, can find the best solution in an optimization problem through forward & backward movements, as well as a step-wise decision process. The parameters used for the application of the ACO optimization technique are summarized in Table 1.

Therefore, identity probability is invoked in setting the initial magnitudes of the variables. The parameters used for the application of the DE optimization technique are summarized in Table 1.

GA algorithm is an evolution-based algorithm that was developed based on Darwin's Principle of Natural Selection for addressing numerous optimization problems (Koza 1994; Yang & Honavar 1998; Maulik & Bandyopadhyay 2000; Deb et al. 2002; Levasseur et al. 2008; Iba & Aranha 2012; Kubat 2017). The algorithm is initialized by generating an initial random population of individual solutions to the considered problem. The goodness of fit for each solution is then assessed using suitable metrics, and crossover and mutation operators are employed to generate the next generation of individual solutions and allow the population to evolve towards an optimal solution. The parameters used for the application of the GA optimization technique are summarized in Table 1.

For live-bed conditions, two Gaussian and two linear functions are defined as the membership functions of the input and output variables, respectively. For clear-water conditions, one Gaussian and one linear function are considered as membership functions of the input and output variables, respectively. For instance, for a combination of three input variables, the total number of parameters to be optimized is 20 (12 antecedent, input-related parameters and 8 consequent, output-related parameters) for live-bed conditions and 10 (6 antecedent parameters and 4 consequent parameters) for clear-water conditions. The number of membership functions employed in this study is selected based on the available number of data for the training phase, and the number of parameters to optimize is less than the number of training data. Table 2 shows an example of optimized parameters for the ANFIS-PSO model with three input variables.

Several combinations of input variables are considered to identify the optimal predictive model for wave-induced pipeline scour depth under live-bed, as well as clear-water conditions. Individually, seven input combinations, called to ⁠, are evaluated (Table 3). In total, 35 different predictive models are assessed, employing different predictive approaches (ANFIS, ANFIS-PSO, ANFIS-ACO, ANFIS-DE and and different input variable combinations (⁠ to ⁠).

To develop the mentioned AI models, laboratory experimental datasets with a total of 69 scour depth observations from four sources are collected (Lucassen 1984; Sumer & Fredsøe 1990; Pu et al. 2001; Mousavi et al. 2009). The datasets are deemed suitable for the present study because of the following: all the datasets were collected for conditions characterized by KC <100, which imply the pipelines were exposed to the wind wave, the pipe surface was hydraulically smooth, and the current induced by waves was perpendicular to the pipe; the experiments were performed under a wide range of Reynolds numbers, reproducing field conditions; and the channel width in all the tests was large enough to neglect the influence of sidewalls.

The analysis is performed separately for clear-water (21 observations) and live-bed (48 observations) scour conditions. The training-testing data are provided based on the 31–17 (in live-bed condition) and 13–8 (in clear-water condition) observations. The data division adopted between training and testing phases is the result of a trial and error search to attain the best performance. Table 4 summarizes the statistical characteristics of the dimensionless parameters e/D, θ, KC and S/D for the overall dataset considered; the range of variation of the parameters is wide enough to obtain robust results.

To evaluate the uncertainty associated with the input variables, for each observed scour depth, the predicted scour depth is computed in the testing phase for a single model but multiple input combinations (⁠ to ⁠). Then, the uncertainty associated with the input variables is quantified using the same approach described above for the uncertainty related to the model structure. Again, this procedure is carried out for both live-bed and clear-water conditions.

This study examines several ANFIS models hybridized with the different nature-inspired algorithms presented above. Each model uses a different tuning process to obtain the appropriate ANFIS model parameters. Hence, the models provide different prediction performances based on their tuning processes. Comparing the performance metrics of the models is a way to assess the impact of their tuning processes on prediction performance. Specifically, to analyze the prediction performance of the mentioned 35 different predictive models (ANFIS, ANFIS-PSO, ANFIS-ACO, ANFIS-DE and for the input combinations to ⁠), the selected prediction performance indices (⁠⁠) are computed for training, and testing phases and live-bed or clear-water scour conditions (Tables 5–9).

Considering only the classic ANFIS model (see Table 5), exhibits the best prediction performance for live-bed conditions (⁠⁠, for training and ⁠, for testing) among all ANFIS prediction models. For the clear-water conditions, no clearly preferable ANFIS model emerges in training and testing phases, although (⁠⁠, for training and ⁠, for testing) shows the best performance. For both the clear-water and live-bed scour conditions, predictions with higher accuracy result from the training phase rather than the testing phase. Overall, the classic ANFIS model is not robust enough to predict wave-induced scour around the pipelines.

Table 6 indicates that the model offers more accurate predictions than the classic ANFIS model and the best prediction performance among all models considered. In particular, shows the best prediction performance for both live-bed conditions (⁠⁠, for training and ⁠, for testing) and clear-water conditions (⁠⁠, for training and ⁠, for testing).

Tables 7–9 (relative to ANFIS-ACO, ANFIS-DE and ANFIS-GA) show that the input variable combination that results in the best prediction performance is different for live-bed and clear-water conditions. The (⁠⁠), and models are the best performing models for live-bed conditions, whereas the ⁠, and ANFIS-GA-M7 models are the best performing models for clear-water conditions.

It must be noted that Tables 5–9 include both error indices (e.g., RMSE and MAE) and similarity indices (e.g., R² and WI). For instance, the ANFIS‐M1 model in live-bed conditions results in 0.072, 0.048, 0.623 and 0.876 for RMSE, MAE, R² and WI, respectively, in the training phase, and 0.077, 0.054, 0.043, 0.552, respectively, in the testing phase (Table 5). It can, therefore, be observed that the prediction performance is reduced in testing phase by 6.94, 12.50, 93.10 and 36.99%, judging respectively from RMSE, MAE, R² and WI. This means that, for the case of the ANFIS-M1 model, the prediction performance based on the error indices is reduced from training to testing phase less significantly than the prediction performance based on the similarity indices. For the ANFIS-M5 and ANFIS-M7 models, instead, the error indices show a more significant prediction performance reduction than the similarity indices. This behavior excludes an issue of overfitting in our approach because overfitting would produce the same pattern of prediction performance reduction for the testing phase in both error and similarity indices. The low prediction performance indices, when observed in the testing phase, reflect instead of the limitations associated with the dataset sample size and the prediction capability of the model considered. For the best predictive model, ANFIS‐PSO‐M7, the prediction performance reduction from training to testing phase is quantified, using the RMSE and MAE indices, by a decrease in performance of 25 and 30.8%, respectively (in live-bed conditions), and a decrease of 65.7 and 75.8%, respectively (in clear-water conditions). Considering the R² and WI indices, the decrease in performance is 15 and 7.2%, respectively, in live-bed conditions, and 1.6 and 0.49%, respectively, in clear-water conditions. Overall, the error metrics (RMSE and MAE) obtained in testing phase indicate a moderate prediction performance reduction (49.3%), while for the similarity metrics (R² and WI) the reduction (6.1%) is negligible.

Our findings indicate that the classic ANFIS model is not accurate in predicting the wave-induced scour depth around pipelines, especially in clear-water conditions (R² = 0.39 for testing phase) for which the models ANFIS‐PSO (R² = 0.98), ANFIS‐ACO (R² = 0.71), and ANFIS‐DE (R² = 0.69) provide a significantly better prediction performance, as a result of the nature-inspired optimization algorithms introduced in this research for the ANFIS model.

The best prediction performance indices for each model in the testing phase is presented in Table 10.

As mentioned, the model with the input variable combination (all the three variables, ⁠, included) is the model resulting in the best prediction performance for both live-bed and clear-water conditions. Furthermore, the obtained values (also shown in Table 10) quantify the significant improvement in prediction provided by all the ANFIS models optimized using the nature-inspired algorithms compared with the classic ANFIS model for both live-bead and clear-water conditions in the testing phase, with the most significant improvement obtained with the ANFIS‐PSO model (⁠ and ⁠, for testing phase).

A visual performance comparison between the different models is provided in the heat map in Figure 3, based on the standardized and performance indices. The RMSE and MAE indices are standardized using the formula ⁠, while the R² and WI indices are standardized using the formula ⁠, where X is the index value. The resulting standardized values are therefore within the range 0 to 1 with the best index value having a standardized value of 1. As mentioned, the model (dark blue column) has the best performance indices for both live-bed and clear-water conditions. In contrast, and (red columns) offer the lowest performance for live-bed and clear-water conditions, respectively.

The prediction performance is also evaluated on two-dimensional scatter plots comparing the simulated and the observed values of scour depth (Figure 4), where the identity (1:1) line is a reference to visualize how close the simulated and observed values are. For live-bed conditions (Figure 4(a)), the points are generally the closest to the 1:1 line and show the most linear pattern (coefficient of determination ⁠), whereas the points corresponding to the other models are noticeably more scattered (⁠⁠). Likewise, for clear-water conditions, the points are the nearest to the 1:1 line with ⁠, whereas most of the points for the other models suggest overestimation of the predicted scour depth.

Models are also comparatively assessed on a Taylor diagram (Figure 5), considering RMSE, R² and normalized standard deviation (Taylor 2001). Again, in the diagram, the model with the best predictions (i.e., the closest points to the points labeled ‘observed’) is the for both live-bed and clear-water scour conditions.

The variability of the measured and predicted scour depth magnitudes is quantified and compared for the different models by computing and plotting the quantiles ⁠, and (Figure 6). The median scour depth predicted by the model is the closest to the median observed value for both live-bed conditions (⁠⁠) and clear-water conditions (⁠⁠). From a comparison of the interquartile range ⁠), which is the difference between Q_75% and Q_25%, it appears that the predictive models underestimate the variability of the observed data in live-bed conditions. In contrast, they either under- or overestimate the measured values in clear-water conditions.

Sharafati et al. (2018) improved Etemad-Shahidi et al. (2011)’s equations using the stochastic GLUE and SUFI approaches and developed Equations (9) and (10). The best proposed model (⁠⁠) was compared with the mentioned formulations from literature using prediction performance indices (Table 11) and several visual performance comparisons (Figure 7).

In the two-dimensional scatter plots (Figure 7(a) and 7(b)) the proposed model points are closer to the 1:1 line and show a more linear pattern (coefficient of determination closer to unity) compared to Etemad-Shahidi et al. (2011)’s and Sharafati et al. (2018)’s models for both live-bed and clear-water scour conditions. The better performance of the model is also confirmed by the heat maps (Figure 7(c) and 7(d)) and the Taylor diagrams (Figure 7(e) and 7(f)).

Although the ANFIS‐PSO model provides a better prediction performance compared to the formulas obtained by Etemad-Shahidi et al. (2011) and Sharafati et al. (2018), its additional structure complexity may potentially hinder its application by some practitioners in the field of scouring. Indeed, the ANFIS model comprises several unknown parameters which needs tuning with an optimization algorithm such as PSO. Hence, we especially advise to use the model proposed in the present study in high-stakes pipeline projects that require a very accurate prediction of scour depth.

To verify that the best predictive model (ANFIS‐PSO‐M7) is consistent with the physics of the pipeline scour phenomenon, Figure 8 shows how the normalized scour depth S/D is predicted to vary with varying e/D, θ and KC in live-bed conditions. The ANFIS‐PSO‐M7 correctly predicts S/D to increase for decreasing e/D or increasing θ and KC, as observed in physical investigations (Sumer & Fredsøe 2002).

The uncertainty associated with the model structure is evaluated considering the considered five models (ANFIS, ANFIS‐PSO, ANFIS‐ACO, ANFIS‐DE and ⁠) with input variable combination (the best performing combination as shown earlier). The uncertainty associated with the input variables is assessed for the model (the best performing model, as explained earlier) and different input variable combinations (⁠⁠). Figures 9 and 10 show the generated 95 PPU band for model structure and input variable uncertainty, respectively. The figures also show the corresponding observed values and are provided for both live-bed and clear-water scour conditions.

From Figure 9, the uncertainty in predicted scour depth associated with the model structure in clear-water (⁠⁠) is higher than in live-bed conditions (⁠⁠). From Figure 10, the uncertainty in predicted scour depth associated with the input variables is also higher in clear water (⁠⁠) compared with live-bed conditions (⁠⁠). It can generally be concluded that prediction of scour depth caused by waves at pipelines in clear-water conditions is characterized by more considerable uncertainty, due to both model structure and input variables, than in live-bed conditions. Furthermore, the uncertainty associated with the input variables is larger than the one associated with the model structure ⁠.

This study proposed and assessed the application of nature-inspired optimization algorithms to enhance the ANFIS model performance in predicting wave-induced pipeline scour depth. The considered algorithms (PSO, ACO, DE and GA) are alternatives to the common trial and error methods for optimization, which are not time-efficient and may lead to overfitting.

The proposed models were trained and tested using four datasets (Lucassen 1984; Sumer & Fredsøe 1990; Pu et al. 2001; Mousavi et al. 2009), considering different combinations of input variables (⁠⁠) derived from dimensional analysis. The prediction accuracy of the various proposed models was assessed based on indices of prediction performance (RMSE, MAE, R², WI) and visual comparison (heatmap of standardized performance metrics, scatter plot, normalized Taylor diagram and boxplot of the predicted and observed scour depth). From the comparison results, it emerged that the ANFIS model including all the three input variables and optimized using a PSO algorithm provides the most accurate wave-induced pipeline scour depth predictions, for both live-bed and clear-water scour conditions.

This paper also evaluated two sources of uncertainty associated with the scour depth prediction, disaggregating the uncertainty of the model structure (type of optimization algorithm) and the one due to combination of the input variables (selection of input variables for the model). To evaluate uncertainty, a Monte Carlo simulation technique is used, and the 95% prediction uncertainty is quantified through the index. From the results, the model structure and the input parameters selection both lead to more considerable uncertainty in scour depth prediction for clear-water conditions than for live-bed conditions. Also, the uncertainty due to the input variable combination is larger than the model structure-associated uncertainty.

This study shows that a relatively simple improvement in the optimization of an ANFIS model, based on the PSO algorithm, may lead to significant improvement in prediction performance not only in comparison with a classic ANFIS model optimized through trial and error procedure but also in comparison with recently developed models based on the MT approach (Etemad-Shahidi et al. 2011) or GLUE and SUFI stochastic approaches (Sharafati et al. 2018). The added complexity of the ANFIS‐PSO model compared to simpler formulations (still suitable for most projects) is counterbalanced by a higher accuracy.

This paper provides new insight into scouring depth prediction for the design of submarine pipelines. Although the use of conventional equations is straightforward for practical purposes, their prediction is not always accurate. This study shows that an ANFIS‐PSO model can be trained and applied for more accurate scour depth predictions that can support a more robust and safer design.

Data cannot be made publicly available; readers should contact the corresponding author for details.