An artificial neural network based system for wave height prediction

We present a system for predicting the hourly significant wave height at a specific wave measurement station in the middle of Israel's Mediterranean coast (Hadera). Our system uses an artificial neural network (ANN) composed of two sub-networks. We evaluate the importance of different inputs to the system. The input includes wind forecast data from the SKIRON atmospheric modeling system, wave forecast for the station's location given by the SWAN wave model, and observed wave data. Our system pre-processes the wind data using a spatial filtering scheme and then enters it into the first sub-network in the form of a multidimensional tensor. We take special care to interconnect the tensor elements through a dimensional permutation that leads the ANN to sum elements along all the tensor's dimensions. Our system groups the output of the first sub-network with the rest of the input and feeds it to the second sub-network that gives the prediction. Our ANN system outperforms the SWAN wave model in estimating wave heights over 1.5 meters. We obtain the best performance when either all input components are used or just wind and observations. Reimplementation of the system at Ashkelon yields smaller improvements due to insufficient training data.

Automated summary

We propose a system for predicting the hourly significant wave height at a specific wave measurement station by utilizing an artificial neural network (ANN) composed of two sub-networks. The system incorporates wind forecast data, wave forecast data, and observed data. The ANN system performs better than the existing SWAN wave model in estimating wave heights over 1.5 meters, showing the importance of all input components or just wind and observations. The system's reimplementation in Ashkelon yields smaller improvements due to insufficient training data.