Implication of General Regression Neural Network (GRNN) For Wave Overtopping Prediction of Vertical Sea Defences

Wave overtopping prediction is one of the essential processes required in designing crest level and other geometrical parameters of coastal structures. In this study, optimisation of the efficiency of a general regression neural network (GRNN) in wave overtopping prediction of vertical structures was considered. Data were selected from wave overtopping database collected by the European research project (EU-CLASH) [1–3], and the dimensionless parameters were chosen based on [1]. The relative crest height, Rc/Hm0,t, was chosen as the discriminator parameter, which greatly increased the accuracy of the model. The same NN configuration was also trained to predict wave overtopping using data from all types of structures. The GRNN-which trained on only vertical data-reduced the number of needed input parameters to be 10 dimensionless parameters. The results from these two NNs for predicting vertical wall cases were compared. The NN accuracy was evaluated using statistical parameters, including correlation coefficient (R), root mean squared error (RMSE) and mean absolute percentage of error (MAPE) values. The GRNN, which was trained on vertical wall data, achieved a high level of accuracy, producing R = 0.9844, RMSE = 0.0014and MAPE= 96.71, with R = 0.9245, RMSE = 0.0054 and MAPE = 62.30 for the one trained on comprehensive data. Although the comprehensive model predicts the vertical wall cases with relatively good accuracy, some extreme errors and dispersion are present. These results illustrate the need for developing NNs on single types of structures. The model was also compared to the EurOtop model [1]for predicting wave overtopping of vertical structures and effectively outweighs it in accuracy. The prediction model will be freely available by the end of this project. Index Terms- Wave overtopping, General Regression Neural Network, Vertical defence structures, CLASH