Ensemble learning based approach for the prediction of monthly significant wave heights

Abstract

The monthly significant wave height is the average of the highest one-third waves (measured from trough to crest) that occur in a month. Accurate prediction of monthly significant wave heights is of great significance to wave power generation, marine traffic, disaster prevention and mitigation. This paper presents a novel stacked ensemble model for the prediction of monthly significant wave heights. 128 sets of data collected from a buoy station offshore the Atlantic Ocean were used to build the proposed models. Firstly, seven artificial intelligence (AI) models, namely the random forest, regression tree, long short-term memory, M5 model tree, adaptive neuro fuzzy inference system, least squares support vector machine optimized by improved particle swarm optimization, and back propagation neural network, were used to predict the monthly significant wave heights. Then, five statistical indices (e.g., coefficient of determination (R²), mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE) and discrepancy ratio (DR)) were used to evaluate the performance of the models. On the basis of the prediction results, three base models with good performance were selected from these seven models, and a novel stacked ensemble model was established to predict the monthly significant wave heights. The results of comparison between the stacked ensemble model and the other three AI base models show that the R², MAPE, MAE and RMSE values of the stacked ensemble model were 0.9426, 3.198 %, 0.0575 m and 0.006 m, respectively, for the training datasets and 0.8564, 6.169 %, 0.100 m and 0.037 m, respectively, for the testing datasets, indicating that the stacked ensemble model has high prediction accuracy for monthly significant wave heights. In addition, the sensitivity and generalization ability of the stacked ensemble model were also analyzed in this study.