Prediction of Significant Wave Height and Investigation of Long-term Trend of Stormy Wave Height based on Long Short-term Memory Neural Network

Abstract

 Deep learning technique can be one of promising options for time-effective wave forecasting system. One of shortcomings of these technique however is that the predictive ability of the model is purely based on the training data, and thus the developed system may not be able to represent the phenomena unexperienced in the past. On the other hand, such shortcomings may be useful to detect unexperienced phenomena: such as the influence of the climate change. To answer this hypothesis, this study applied long short-term memory (LSTM) recurrent neural network technique, and developed a model for forecasting of significant wave heights as functions of the recent history of local wave heights and wind speed. The model was trained based on the data observed from the year 2000 to 2007, and the trained model was tested against the data observed after 2008. While the model showed reasonably good predictive skills of the wave height, the model tended to underestimate the extreme wave heights and this underestimation showed increasing trend in the recent years. Similar long-term trend of underestimation was also found in the error of wind speed. Underestimations of wind and wave showed linear relationships, which were consistent with empirical Wilson's formula, meaning that amplification of extreme wave height is mostly due to intensification of the wind speed.