Closure to “Predictions of Missing Wave Data by Recurrent Neuronets” by Can Elmar Balas, Levent Koç, and Lale Balas

Abstract

The authors compared different methods for estimating missing wave data recorded by directional wave buoys. They developed AR models, ARX models, multilayer feed forward MFF models and recurrent Elman-type neural network ENN models trained with the learning algorithms SDM and CG. Only data from 12 months of noncontinuous measurements at the directional wave buoy of Alanya Turkey were used for training the models. In the writers’ opinion, the less than two weeks of data used for training the models is an unacceptably low volume of wave-climate information to properly characterize any wave measurement station. The data set used for training is shown in Table 1; 147 data points corresponding to 292 h obtained from the Alanya directional wave buoy and meteorological station at 2 h intervals were used. Tables 3 and 7 show the neural network structures of the MFF and ENN models proposed by the authors to characterize the Alanya wave time series of significant wave height Hs t , mean period T t , and wave direction WD t . One time series of 147 data points less than 2 weeks of wave and wind information was used by the authors to train models as x3h15y3 108 free parameters and x12h35y3 563 free parameters . Overlearning is likely to take place because the number of free parameters of the proposed models is similar to or higher than the number of data points used for training. Furthermore, a proper characterization of the wave climate at any measurement station using information of less than 2 weeks in a year should never be expected, because there are seasonal differences i.e., winter, spring, and summer . Whatever the wave-climate model to be proposed was, the supporting wave-climate information for training or calibrating models should be at least a full year to cover the most common wave-climate situations during the year. The models described by the authors are based on only a few days of continuous records, and Table 1 shows significant differences with the short periods of wave climate information used for validation and testing. The dimensional mean absolute error ME and the dimensionless determination coefficient r shown in Tables 4, 5, 8, and 9 are not consistent. For instance, in Table 4, Test 1, ENN model, wave direction WD ° shows a very low absolute error, ME=0.0169°, while coefficient r=−0.01 is very poor. Not only