ARTICLE IN PRESS Wave tranquility studies using neural networks S.N. Londhea, M.C. Deob,* b a Sinhgad College of Engineering, Pune 411041, India Indian Institute of Technology, Civil Engineering, Mumbai 400076, India Abstract Information on heights of waves and their distribution around harbor entrances is traditionally obtained from the knowledge of incident wave, seabed and harbor characteristics by using experimental as well as numerical models. This paper presents an alternative to these techniques based on the computational tool of neural networks. Modular networks were developed in order to estimate wave heights in and around a dredged approach channel leading to harbor entrance. The data involved pertained to two harbor locations in India. The training of networks was done using a numerical model, which solved the mild slope equation. Test of the network with several alternative error criteria confirmed capability of the neural network approach to perform the wave tranquility studies. A variety of learning schemes and search routines were employed so as to select the best possible training to the network. Mutual comparison between these showed that the scaled conjugate method was the fastest among all whereas the one step secant scheme was the most memory efficient. The Brent’s search and the golden section search routines forming part of the conjugate gradient Fletcher–Reeves update approach of training took the least amount of time to train the network per epoch. Calibration of the neural network with both mean square as well as the sum squared error as performance functions yielded satisfactory results. Keywords: Neural networks; Harbor tranquility; Training algorithms; Numerical wave models; Neural network performance 1. Introduction An essential feature of harbor planning is wave tranquility studies in and around a harbor. When waves advance from deep to shallow water their heights and attacking ARTICLE IN PRESS 420 angles change because of the effects of shoaling, refraction, diffraction, breaking and reflection. Many harbors involve an approach channel dredged to a certain depth. This is necessitated from the catering to large ships having deeper drafts for navigation. When waves from open sea invade in and around such channels they are subjected to abrupt changes in topography and combined action of refraction and bathymetry-induced diffraction. This can completely change the wave height distribution around the harbor causing attenuation at some places and concentration of waves at some other locations. Information on heights of waves and their distribution around the harbor entrance is obtained from the knowledge of incident wave and seabed characteristics by using both experimental modeling as well as numerical modeling. Experimental or physical models require considerable amount of time and may not be cost effective. They also suffer from (physical) laboratory effects such as scale effects. Because of the Froudian scaling involved they often fail to account properly for wave reflection and dissipation effects which could be dominated by surface tension and viscosity. However hydraulic modeling may become unavoidable when complex site conditions are concerned. Numerical models provide approximate solution of a governing differential equation describing the wave propagation. The governing equation could be either of mild slope or of Boussinesq type. Berkhoff’s [1] mild-slope equation was presented to solve for linear wave attack under steady state or time-dependent conditions. Finite difference as well as finite element schemes have been used by different investigators in order to provide for various approximations needed to solve the mild-slope equations [2]. Available solutions range from Berkhoff’s [1] twodimentional combined refraction–diffraction equation of elliptic form to Copeland’s [3] hyperbolic and Darlymple and Kirby’s [4] parabolic form. Application of the mild slope equations is relatively straight but it involves less satisfactory modeling of the flow physics caused by their underlying assumptions. Another class of numerical models is therefore alternatively followed and this is based on the Boussinesq equations. The Boussinesq equations represent a set of simultaneous, non-linear, hyperbolic equations and contain all important wave processes, such as shoaling, refraction, diffraction and reflection. Various modified forms of the original Boussinesq equations have resulted in accounting for unsteady time-dependent conditions, dispersion characteristics or wave directionality effects as well as wave non-linearity in the solution procedure. This has led to increased usefulness of the numerical schemes in a wide range of water depths unlike the earlier attempts restricted to shallow water regions. Latest information on the physical and numerical studies can be seen in Young [5] and Edge and Hemsley [6]. Considerable progress has been made in numerical modeling while simulating as many real sea conditions as possible like higher order wave theory, complex boundaries and presence of current. However inherent computational difficulties like rounding off and truncation errors, speed and memory requirement coupled with complexity of the phenomenon involved in such a modeling may not always yield highly satisfactory results. Efforts to tackle the problem of wave propagation into a S.N. Londhe, M.C. Deo / Marine Structures 16 (2003) 419–436 421 Fig. 1. Location plan. harbor using new schemes should therefore be welcomed [7]. Artificial neural networks (ANN) or simply neural networks (NN) could be one such technique. ANN basically provides mapping between random inputs and outputs irrespectively having knowledge of the underlying phenomenon. Unlike traditional analysis tools it can learn from the examples or training patterns fed to it, store the knowledge in its weights and bias values and use them to estimate future values. Although their use to solve computational problems in civil engineering started since last two decades, applications of ANNs to tackle ocean engineering contexts are too scarce and there exists a great room for attempt in this regard. The aim of this work was to employ the technique of ANN to study wave attenuation along a dredged approach channel leading to a harbor entrance and compare the results so obtained with numerical and physical model outputs. The study also provides analysis of network performances using a variety of training algorithms, search routines, error goals and error criteria. The present work is based on data available at two harbor locations, namely New Mangalore Port and Bina Bay Harbor near Goa located along the western Indian coastline (see Fig. 1 for location of these sites). 2. Study area The studies have been carried out with the help of two data sets pertaining to locations of New Mangalore Port and Bina Bay (Marmugao) port along the west coast of India (Fig. 1). The available information at the former location was more exhaustive and hence major studies were concentrated on this data set. As mentioned above the New-Mangalore port is situated on the west coast of India and its harbor is connected to the sea by an approach channel which is 5200 m long, 245 m wide and dredged to –13.5 m below the chart datum with side slopes of 1:10 (Fig. 2). The ARTICLE IN PRESS 422 Fig. 2. New Mangalore Harbor (Ref. [10]). (Dots represent the grid 250 m 50 m from seaward boundary to harbor entrance.) channel is oriented at 259 with respect to North, so as to make it in-line with predominant wave direction in the region to achieve maximum wave tranquility in the harbor. The developed neural network had inputs in the form of wave heights, wave periods and angles of wave attack at 9 grid points along a cross section across the approach channel. The output from the network was the changed wave heights at each of the 9 grid points located at the subsequent cross section located 50 m on downstream side. The input–output representation made in this way was adopted for convenience and according to availability of data for network training in that manner. The output of a numerical model REFDIF1 was used in the present studies to train the neural network. The model involves numerical solution of the parabolic approximation of mild slope equation, and it explicitly simulates combined effects of refraction and bathymetric diffraction over an approach channel. This was in the form of attenuated wave heights at 9 grid points along different cross sections of the channel, spaced 50 m center to center corresponding to an incident wave height of 3.66 m. 3. Modular neural networks A network architecture called modular network consisting of several modules with limited interconnections between them was employed. Its use was found necessary because initial trials had shown that the network training was extremely slow or ARTICLE IN PRESS 423 unattainable due to a large number of training patterns involved. Modularity permits separate tackling of smaller tasks using small modules and then combination of these modules in series or parallel form [8]. This helps formation of a network for the specific task from a set of previously trained modules. Also the system can be expanded without redeveloping the whole package afresh in order to include new modules when such a need arises [9]. This approach was found to be most suitable for designing NN for the present problem because there were as many as 4 input vectors, each containing 9 values and 1 output vector having 9 values. Ninety-five such vectors made a total of 4275 values as per the data set. 4. NN models The modular NN used in the present study had either 3 or 4 modules in the input layer. Each module pertained to one input vector. Fig. 3 indicates the 4-module network that was employed. The 3-module network had wave height, wave period and seabed levels as input vectors while the other one had the angle of wave attack as an additional input vector. The output of both networks was heights of the wave at the subsequent cross section. Development of such separate networks was guided by applications of the numerical model made in advance and by availability of data in that format. (Input 1) [9x1] Weights[9x9] weights[9x9] ∑ (Input 2) [9x1] Transfer Function weights[9x9] bias [9x1] (Input 3) [9x1] weights[9x9] (Input 4) [9x1] Fig. 3. 4-Module network. output [9x1] ARTICLE IN PRESS 424 1.8 network output target 1.6 wave height (m) 1.4 1.2 1 0.8 0.6 0.4 0.2 500 1000 1500 2000 2500 3000 3500 4000 distance from seaward boundary (m) 4500 5000 Fig. 4. Testing performance (ANN with lab data). A better way of imparting training to the network would have involved the use of physical model output for this purpose. The physical model covering the entire length of 5200 m of the approach channel was constructed to a geometrically similar scale of 1:150. The model extended upto 13.0 m contour with respect to Chart Datum on the seaside. The bathymetry extending upto 1200 m on either side of centerline of the channel was reproduced in order to simulate the transfer of wave energy across the wave crest properly. However it is extremely difficult to collect simultaneous observations of wave heights at several locations as per the training requirement. None the less hydraulic model experiments reported in Kanetkar et al. [10] were used as a trial to train the network. The model measurements were available at the centerline of the approach channel corresponding to incident wave height of 3.66 m and wave period of 9 s. Fig. 4 shows the results of network testing. It indicates how the wave height attenuated along the channel axis when physical as well as numerical model based training was employed. The close proximity between the results obtained through the two schemes (numerical and ANN) showed that one can train the network using numerical model output as well. It is expected that owing to the high degrees of freedom involved, the network could learn information not only as contained in the numerical model but also something over and above that. 5. Training algorithms A variety of training schemes was employed to impart training to the network. This was done to make sure that adequate training is imparted and also to study relative performance of each and converge to the best scheme. The ASCE Task committee [11] had earlier appreciated the need to actually compare the network performance by various algorithms. Thirumalaiah and Deo [12] had found that network performance as well as prediction accuracy can be improved by appropriate S.N. Londhe, M.C. Deo / Marine Structures 16 (2003) 419–436 425 choice of training. The methods involved were: (i) gradient descent with momentum (GDM), (ii) resilient backpropagation (RP), (iii) conjugate gradient Fletcher–Reeves update (CGF), (iv) conjugate gradient Polak–Ribiere update (CGP), (v) Powell Beale restarts (CGB), (vi) scaled conjugate gradient (SCG), (vii) Broyden, Fletcher, Goldfarb and Shanno update (BFG), (viii) one step secant algorithm (OSS) and (ix) Levenberg–Marquardt algorithm (LM). The details of these algorithms can be found in Demuth et al. [13]. GDM has provision for faster convergence while RP is aimed at eliminating problems arising out of smaller magnitudes of the error gradients. A general conjugate gradient scheme involves performing a search along conjugate direction in order to determine step size to minimize performance function. Five search techniques are alternatively used in the current study. They are (i) golden section search, (ii) Brent’s search, (iii) hybrid bisection cubic search, (iv) Charalambus search and (v) backtracking. There are four different versions of the conjugate algorithm (as per the technique of finding a new search direction) namely CGF, CGP, CGB and SCG. The BFG scheme is based on the Newton’s method but does not require calculation of the second derivatives of the error gradient. The BFG scheme updates an approximate Hessian matrix at each iteration of the algorithm. OSS is a secant approximation falling in between the quasi-Newton’s method and the conjugate gradient algorithm which assumes at each iteration that the previous Hessian matrix was the identity matrix. The LM algorithm is designed to approach second-order-training speed without having to compute the Hessian matrix. 6. Testing results The first network was trained on the basis of data belonging to a constant wave period of 11 s. It was then tested for unseen inputs for the case of constant wave periods of 7 and 9 s, respectively. Fig. 5 shows how the network performed during testing with respect to the case of input wave period of 9 s. It indicates how the wave 2 wave height (m) neural network numerical model physical model 1.5 1 0.5 0 0 1000 2000 3000 4000 5000 distance from seaword boundary (m) Fig. 5. Comparison of numerical, physical and ANN model. ARTICLE IN PRESS 426 height attenuated along the central axis of the channel. The reduction in wave height is due to the reduction in propagating wave energy caused by refraction away from the channel. Fig. 5 also shows network output of wave attenuation along the harbor channel vis-a" -vis the same yielded by numerical as well as physical model. It may be noted that the network output agrees very well with the numerical as well as the physical model outputs. During the testing, the wave height, period and sea bed elevations at the seaward boundary along the 9 grid points of the first cross section were provided as input and output wave heights at the next cross section were obtained using the weights and biases fixed at the end of training. These wave heights along with the seabed elevations at that section were then used as input for obtaining wave heights at the next cross section. This process was continued till the harbor entrance was reached. The mean square error goal for every algorithm was decided by making sensitivity analysis involving selection of different goals for training and then comparing the testing results with those of the numerical model. It may be noted that input and output wave heights are scaled down to half in accordance with the available data (1.83 m instead of 3.66 m). The training and testing was carried out on a Pentium 2 processor with 64 MB RAM memory which took 2.36 s for the testing exercise. Where as the numerical model was run on a silicon graphics work station with which took about 60 s time. All the 9 algorithms gave satisfactory results, but the results of CGF with Charalambous’ search as the default line search routine were the best as indicated by the scatter plots and accompanying high values of correlation coefficient of 0.99 (see Fig. 6 corresponding to testing with respect to 7 s) along with achievement of a low mean square error goal. When the model was tested with other algorithms the values of correlation coefficients were not at par with those obtained using the CGF except the algorithm of BFG. But the CGF algorithm had an added advantage that it was memory efficient than BFG in that it required lesser time per epoch than the BFG. 2 target wave height (m) 1.8 1.6 1.4 1.2 best fit Network output 1 0.8 0.6 0.4 0.2 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 actual wave height (m) Fig. 6. Actual against target wave heights (t ¼ 7 s) 2 427 2200 2 2 1.5 2 2.5 2000 1 2 Lateral Distance (m) 2 1800 1 1600 1400 1200 1000 800 0.5 2 1.5 600 400 2 500 1.51.5 1 0.5 0.5 1000 1500 2000 2500 3000 3500 4000 4500 distance from seaward boundary (m) Fig. 7. Wave contour plot (t ¼ 7 s). The wave contour plot for a typical input wave period of 7 s is shown in Fig. 7, which clearly indicate the wave attenuation along centerline of the approach channel and also predict the wave heights at any desired location in the model area. Note that all contours presented are in meter. When the incident wave direction is parallel to the channel axis (as underlying this figures), waves at the two sides will be higher than those inside the channel since wave celerity is higher in the channel (because of deeper water). This supported the observation of Li et al. [14] reported in an earlier study. 6.1. Effect of change in angle of attack The study reported in the preceding paragraphs was carried out on the basis of the 3-module network. Thereafter the angle of wave attack, y; made by the wave with the channel axis was taken as an additional fourth input and the corresponding 4module model was run. The numerical model data were available for a few values of y ranging from 5 to 35 . However attenuation effect was significant only up to y ¼ 15 : In the present study the network was trained for y ¼ 10 : Fig. 8 shows wave attenuation along the harbor channel when y changes from 5 to 15 . Note that all contour presented are in meter. The final wave heights at the harbor entrance do not appear to change much indicating lesser refraction effects for higher angles. This trend is almost similar to the one obtained by the numerical model. 6.2. Effect of initial wave height on wave attenuation Harbor planning exercise often calls for studying changes in the wave distribution with respect to small variations in design incident wave heights. The 3-input model was originally trained for an incident wave height of 3.66 m (at the seaward boundary) along with a wave period of 11 s. It was now run for different initial wave heights, retaining the weights and biases of the trained network. This indicated an ARTICLE IN PRESS 428 2 wave height (m) 1.5 testing 7 sec. testing 9 sec. 1 training 11 sec. 0.5 0 0 1000 2000 3000 4000 5000 distance from seaward boundary (m) Fig. 8. Effect of oblique wave incidence on wave attenuation. Table 1 Effect of change in initial wave heights Initial wave height (m) (1) Final wave height (m) (2) % attenuation (3) 1.50 1.70 1.75 1.90 0.24 0.14 0.14 0.05 84.28 91.61 93.18 97.40 Note: Wave heights are scaled down to half. increase in the percentage wave attenuation near the harbor entrance with an increase in initial wave height as shown in Table 1. As the initial wave height increases more attenuation at the harbor entrance is noticed because of increased dissipation for higher waves. 6.3. Effect of change in wave period The effect of change in the wave period on wave attenuation along the harbor channel as given by the network is shown in Fig. 9, which is exactly the same as produced by the numerical model in earlier studies. Larger wave periods result in lower levels of wave heights along the channel (or higher attenuation) because of the increase of wavelength and the reduction in the steepness ratio, leading to higher attenuation of flatter waves. 6.4. Network performance analysis Relative performances of different training algorithms were drawn in Table 2 (for 3-input network) which give the number of epochs, mean square error (mse) reached, r7 ; r9 (which are correlation coefficients corresponding to testing with 7 and 9 s 429 2 wave height (m) 1.5 testing 7 sec. testing 9 sec. 1 training 11 sec. 0.5 0 0 1000 2000 3000 4000 5000 distance from seaward boundary (m) Fig. 9. Effect of change in wave period on Wave attenuation (ANN approach). Table 2 Test results for wave periods of 7 and 9 s (training wave period=11 s) Algorithm (1) Epochs (2) mse (3) r7 (4) r9 (5) GDM CGF CGP CGB SCG LM BFG OSS RP 18232 520 175 13 50 1 28 2072 2500 0.0099 0.0044 0.0099 0.0099 0.0043 0.0040 0.0097 0.0042 0.0056 0.98 0.99 0.98 0.95 0.95 0.97 0.99 0.98 0.95 0.99 0.99 0.99 0.95 0.98 0.98 0.99 0.99 0.97 Note: r7 ; r9 =correlation coefficient for wave period 7 and 9 s, respectively; Epochs=iterations or pa period, respectively). A highly satisfactory performance by all algorithms may be noted. The LM algorithm converged in just one epoch while the GDM converged to amounts of 18,232 epochs. 6.4.1. Cross-validation Optimum learning as well as performance of the network can be achieved by crossvalidation where an independent data set is used to assess its performance during various stages of learning. Cross-validation also results in avoiding overfitting errors, usually when very large number of training pairs are supplied, as in the present case. Both the network models mentioned earlier were therefore subjected to crossvalidation. In case of the 3-module model, the network was trained with the help of data corresponding to a wave period of 11 s and validated and tested using data of 9 and 7 s, respectively. The 4-module model was trained, validated and tested for data sets belonging to angles of 10 , 5 and 15 , respectively. The results indicated that in general the mse ARTICLE IN PRESS 430 reduced in respect of the training as well as the validation data set during different stages of training (increasing epochs) shown for example in Table 3 which additionally shows absence of any overfitting. 6.4.2. Efficiency of algorithms In the light of various options of network training it was thought to be of help if some guidance on the algorithm selection is given to the user. The efficiency of an algorithm will depend on many factors including the complexity of the problem, the number of data points in the training set, the number of weights and biases in the network, the error goal and whether the network is used for pattern recognition (discriminant analysis) or function approximation (regression). Tables 4 and 5 summarize the results out of the seven training methods of the network such as algorithms of CGF, CGB, CGP, SCG, LM, BFG and OSS. Each entry in the tables is a result of 25 different trials, each involving different random initial weights. The mean time required for each training exercise and their ratio with respect to the lowest time (columns 2 and 3, Table 4) indicates the speed of the algorithm. Standard deviation of the time required for training (with respect to mean time) as well as minimum time and maximum time required for training are also indicated in columns 4–6 of Table 4. In each case, the network training was Table 3 Cross-validation for the 3-input network (algorithm: CGF) Epochs (1) mse on training set (wave period 11 s) (2) mse on validation set (wave period 9 s) (3) 000 100 200 300 400 520 convergence 0.3914 0.0078 0.0058 0.0050 0.0047 0.0044 0.2453 5.59 105 0.1600 0.1191 0.0728 0.0539 Table 4 Speed comparison Algorithm (1) Mean time (s) (2) Ratio of mean times (3) Min. time (s) Max. Time (4) (s) (5) Std. deviation (6) Corr. coeff. (7) CGF CGP CGB SCG LM BFG OSS 25.56 68.90 10.98 2.49 5.54 30.03 103.18 10.25 27.63 4.40 1.00 2.22 12.04 41.37 14.72 31.37 4.78 2.31 5.39 28.01 76.24 5.380 21.000 3.510 0.428 0.460 1.760 13.980 0.99 0.99 0.99 0.99 0.99 0.99 0.99 34.71 111.55 17.69 4.56 7.80 33.39 123.09 ARTICLE IN PRESS 431 Table 5 Efficiency of algorithms Algorithm (1) Mean epochs (2) Mean time (s) (3) Time per epoch (s) (4) CGF CGP CGB SCG LM BFG OSS 372 1129 140 39 1 71 1848 25.56 68.90 10.98 2.49 5.54 30.03 103.18 0.068 0.061 0.078 0.063 5.540 0.420 0.056 continued till mse was reduced to a low figure of 0.0045. It can also be seen that the SCG algorithm required the minimum time to train, though the LM algorithm reached the error goal in 1 epoch. The next fastest algorithm was the CGB, which was however 4.5 times slower than the SCG. The CGF and BFG algorithms performed equally in terms of time required for convergence whereas the CGP and OSS algorithms were much slower; the slowest scheme was the OSS. The fastest algorithm for this particular problem was therefore the scaled conjugate gradient algorithm (SCG). Its speed is possibly due to the fact that it avoids the time consuming line search, which was more than two times faster compared with the next fastest algorithm LM as indicated in Table 4, although the time required per epoch was the least in case of the OSS indicating it as the most memory efficient as shown in Table 5, which gives mean number of epochs, mean time to reach the error goal and the time taken per epoch for various training schemes. All algorithms performed equally well as far as accuracy in the simulation of training data set was considered as indicated by the high value of the correlation coefficient, ‘r’ (=0.99). However it is noteworthy that this trend may not continue in actual application because one additional epoch may adjust the weights in such a manner that networks may overfit yielding erroneous results. Hence a sensitivity analysis is recommended to fix the goal, for which the accuracy of different training algorithms may not be equivalent so far as the testing is considered. 6.4.3. Model assessment The work discussed so far involved use of the performance function of mse to achieve the error goal with correlation coefficient as an indication of the model performance during testing. Although the criterion of correlation coefficient is very common in assessing model effectiveness it does not identify specific regions where the model is deficient. Other error measures were therefore employed to study the model performance. These are average error (AE), mean absolute error (MAE), mean relative error (MRE), root mean square error (RMSE), mean squared relative error (MSRE), the coefficient of efficiency (CE), and coefficient of determination ARTICLE IN PRESS 432 Table 6 Model assessment Location (1) Testing condition (2) AE (3) MAE (4) MRE (5) RMS (6) MSRE (7) CE (8) r2 (9) New-Mangalore Port T ¼ 7s T ¼ 9s y ¼ 5 y ¼ 15 T ¼ 10 s y ¼ 14 2.62 2.05 9.42 33.13 0.063 0.058 0.11 0.28 7.85 7.91 13.25 33.46 0.076 0.075 0.142 0.325 0.0099 0.0103 0.0273 0.1553 0.98 0.98 0.91 0.358 0.98 0.98 0.96 0.91 1.80 0.047 2.12 0.057 0.0006 0.01 0.93 Bina Bay (Marmugao) Note: T=wave period; y=angle of approach. (r2 ). The expressions for these error measures are available in Karunanithi et al. [15] and Dawson [16]. The measures of AE, MAE and MRE give different numerical values with respect to the original data. However AE gives algebraic deviations while MAE and MRE present absolute differences disregarding over or under estimation. The difference between MAE and MRE lies in the non-dimensional evaluation of the latter, just like the one between RMSE and MSRE, where, in addition, the distinction between over or under estimation is neglected by taking square of the differences. Relative errors give a more balanced criterion. The lower the values of these measures, the greater is the accuracy. CE and r2 do not depend on data scale and hence are more suited when different scales are involved. CE indicates prediction capabilities of values different from the mean and varies from N to +1. CE of 0.9 and above is very satisfactory and that of below 0.8 is unsatisfactory; r2 varies in the range of (1, 1) and measures the model-explained variance. Table 6 shows the values of these error measures during testing of the models. It involves analysis of data pertaining to New Mangalore as well as Bina Bay harbor. (The latter is discussed in the subsequent section.) Both neural network models— with 3- and 4-input modules were considered. The first two rows in Table 6 pertain to test results with 7 and 9 s wave period (for the 3-module model) while the third and fourth rows belong to testing with 5 and 15 approach angle (for the 4-module model). The testing results for the former model can be seen as the best because it involved lesser errors compared to the latter model. It may be seen that all error measures collectively indicate lower levels of the discrepancy between the ANN model and the numerical model. The neural network models developed in the present work used the performance function of mean square error or ‘mse’. The network performance with respect to another function of sum squared error or ‘sse’ was also investigated. Table 7 shows the result. It gives the number of epochs required to achieve the error goal, the time taken to do so, the SSE involved, as well as the correlation for the two testing cases under each model. The high values of the correlation coefficients show that both the models worked satisfactorily for two performance functions namely the ‘mse’ discussed earlier and the ‘sse’. ARTICLE IN PRESS 433 Table 7 Results of using performance function SSE Model (1) Epochs (2) Time (3) SSE (4) r7 (5) r9 (6) r5 (7) r15 (8) 3-input 4-input 428 106 32.13 9.33 4.3998 4.9069 0.99 — 0.99 — — 0.99 — 0.97 Algorithm: CGF, r7 ; r9 =correlation coefficient for testing with respect to wave period of 7 and 9 s; r5 ; r15 =Correlation coefficient for testing with respect to angle of 5 and 15 . Table 8 Effect of Line search routine (Algorithm: CGF; 3-input network) Line search routine (1) mse (2) Epochs (3) Time (s) (4) r7 (5) r9 (6) Time/epoch (s) (7) Golden section Brent’s search Hybrid bisection-cubic search Charalambous’ search Back tracking 0.0066 0.0081 0.0078 0.0044 0.0049 427 127 157 520 915 23.73 7.14 16.2 34.49 7.74 0.98 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.056 0.056 0.100 0.066 0.059 Note: r7 ; r9 =correlation coefficient for wave period 7 and 9 s, respectively. 6.5. Effect of line search routines on network performance Many variations of the conjugate gradient and quasi-Newton algorithms require that a line search be performed to determine the step size, which will minimize the performance function along that line. Five different search functions were used in the present work namely, Golden Section Search, Brent’s Search, Hybrid BisectionCubic Search, Charalambous’ Search and Back tracking. Any of these search functions can be used interchangeably with all varieties of the conjugate gradient algorithm. Some search functions are best suited to certain algorithms, although the optimum choice can vary according to the specific application. Table 8 corresponding to the 3-module network indicates that all search routines performed equally well as reflected in high values of the correlation coefficients. The Hybrid Bisection-Cubic Search did more computations for every epoch compared to the Golden Section Search or Brent’s Search as indicated by its time consumption per epoch. For the present problem all the search routines performed more or less in equal efficiency as seen by their relative speed and accuracy. 6.6. Wave attenuation at Bina Bay Port (Marmugao harbor) The discussion presented under previous sections so far is based on results of data analysis carried out for the site of New Mangalore port. In order to examine portability of the developed (trained) network to any another location, the data available at Bina Bay near Marmugao harbor was analyzed. This is described below. ARTICLE IN PRESS 434 Fig. 10. Study area (Bina Bay)—the computational grids have size 10 m 10 m. 6.6.1. Study area The Bina Bay is located about 5 km south of the existing Marmugao Port along the west coast of India (Refer Fig. 1). The harbor proposed at this location involves an approach channel oriented at 256 N in the predominant wave direction. The proposed channel is 275 m wide, to be dredged to 14.4 m below the chart datum. The study area is shown in Fig. 10. 6.6.2. Data Agarwal et al. [17] studied the phenomenon of wave attenuation in the Bina bay using a numerical model MIKE21-BW developed by Danish Hydraulic Institute, Denmark. This model involves a different governing equation of Boussinesq type unlike the earlier numerical model, REFDIF1, that was based on the equation of mild slope. The numerical model solves the Boussinesq equations after simplification for water surface elevation at each grid point in the time domain. Hence waveforms can be simulated with these models. The MIKE 21-BW model was applied within an area of 3500 m 5000 m. The wave height attenuation at the center line of the approach channel was plotted through out a distance of 1000 m from the seaward boundary to harbor entrance located at 2000 m from the sea ward boundary with an initial wave of height of 2.6 m and period of 10 s. 6.6.3. Training and testing Considering availability of data the 4-module model was adopted. The network trained as in the earlier studies was utilized for this application. The initial input ARTICLE IN PRESS 435 2.6 ANN model Target 2.5 wave height (m) 2.4 2.3 2.2 2.1 2 1.9 1.8 1000 1200 1400 1600 1800 2000 2200 distance from seaward boundary (m) Fig. 11. Wave attenuation using neural network model (Bina Bay Port, Mormugao). consisted of wave height of 2.6 m, wave period of 10 s, angle of incidence of 14 as well as depths of sea bed at 9 grid points along the cross-section at starting point of the seaward boundary. Values of wave heights at the 9 grid points along the subsequent cross-section located 50 m downstream were the output. The process was repeated for 1000 m distance, using the preceding cross-section’s output. Wave attenuation along the centerline of approach channel is shown in Fig. 11, which follows the same trend as that of the numerical model. The model performance was also assessed using the various error measures as explained in earlier section. In this case also all results indicated very less discrepancy between the ANN model and the numerical model outputs as seen in the last row of Table 6 referred to earlier. All the above results indicate that that the developed neural network model worked satisfactorily for this harbor also. 7. Conclusions The foregoing sections presented the development of Artificial neural networks (ANN) models in order to obtain distribution of attenuated wave pattern at the entrance of a harbor involving dredged approach channel. The developed networks were found to follow the expected trend of wave height attenuation along the approach channel of the selected harbors. The trained network when tested for unseen inputs yielded satisfactory output of wave heights, which compared well with the numerical as well as the physical model results. This was confirmed by the scatter plots and accompanying values of a variety of error measures. The technique of ANN thus can be used as an option to the conventional numerical scheme with the advantage that it does not require as much elaborate data as the numerical schemes, it is relatively less complex in both modeling and computations and also less rigid. Application of the ANN model shows that larger incident wave heights undergo higher attenuation as they propagate towards the harbor entrance. Similarly longer incident wave periods indicate increased attenuation. This trend matches with the ARTICLE IN PRESS 436 one observed by earlier investigators while using numerical and physical models. The effect of variation in the incident wave height, period and angles of wave incidence was properly simulated by the neural network. The decrease in error for validation set of data with corresponding decrease in error for training set of data seen during cross-validation showed that the network was devoid of overfitting errors. It was found that the fastest algorithm for the present problem was scaled conjugate gradient algorithm. It required the least time to train whereas one step secant algorithm was most memory efficient in that it required least time per epoch. The network training based on the Conjugate Gradient Fletcher–Reeves update scheme was effected using five different line search routines in order to compare their performance. All the search routines performed satisfactorily. The Golden section and Brent’s search took the least amount of time per epoch indicating their memory efficient nature for the two data sets involved. Calibration of the model with the help of both ‘mse’ and ‘sse’ as performance functions yielded satisfactory results. References [1] Berkhoff JCW. Comparison of combined refraction diffraction. Proceedings of the 13th International Conference on Coastal Engineering. ASCE: Vancouver, Canada, 1972. p. 471–90. [2] Panchang VG, Pearce BR, Ge W, Cushman-Roisin B. Solution to the mild slope wave problem by iteration. Elsevier J Appl Ocean Res 1991;13(4):187–99. [3] Copeland GJM. A practical alternative to the mild slope wave equation. Elsevier J Coastal Eng 1985;9:125–49. [4] Dalrymple RA. Shoring up coastal engineering. ASCE J Civil Eng 2001:71(3):52–3, 84. [5] Young IR. Wind generated ocean waves. Amsterdam: Elsevier Sciences, 1999. [6] Edge BL, Hemsley JM. Ocean wave measurement and analysis. Reston, VA: ASCE, 2001. [7] Dalrymple RA, Kirby JT. Combined refraction/diffraction model REFDIF1. University of Dealware, CACR Report No. 91–2, 1991. [8] Mehrotra K, Mohan CK, Ranka S. Elements of artificial neural networks. Cambridge, MA, USA: The MIT Press; 1997. [9] Flood I, Kartam N. Neural networks in civil engineering 1: principles and understanding. ASCE J Comput Civil Eng 1994;8(2):131–48. [10] Kanetkar CN, Joshi VB, Agarwal JD. Wave propagation in a navigation channel to Harbor. Proceedings of the Indian National Conference on Harbor and Ocean Engineering, 1994. p. J175–84. [11] The ASCE Task Committee. Artificial neural networks in Hydrology II: hydrologic applications. ASCE J Hydrol Eng 2000;5(2):124–37. [12] Thirumalaiah K, Deo MC. River stage forecasting using artificial neural networks. ASCE J Hydrol Eng 1998;3(1):26–32. [13] Demuth H, Beale M, Hagen M. Neural network toolbox user’s guide. The Mathworks Inc., Natick, MA, USA, 1998. [14] Li YS, Liu SX, Wai OWH, Yu YX. Wave concentration by a navigation channel. Appl Ocean Res 2000;22:199–213. [15] Karunanithi N, Grenney WJ, Whitley D, Bovee K. Neural networks for river flow predictions. ASCE J Comput Civil Eng 1994;8(2):201–19. [16] Dawson CW, Wilby RL. Hydrological modeling using artificial neural networks. Prog Phys Geogr 2001;25(1):80–108. [17] Agarwal JD, Ranganath IR, Kanetkar CN. Optimisation of harbour layout using mathematical models. Proceedings of the International Conference on Ocean Engineering (ICOE), IIT Madras, 2001. p. 273–8.
© Copyright 2024 Paperzz