Search

Deep learning methods have excellent prospects for application in wave forecasting research. This study employed the convolutional LSTM (ConvLSTM) algorithm to predict the South China Sea (SCS) significant wave height (SWH). Three prediction models were established to investigate the influences of setting different parameters and using multiple training data on the forecasting effects. Compared with the SWH data from the China–France Ocean Satellite (CFOSAT), the SWH of WAVEWATCH III (WWIII) from the pacific islands ocean observing system are accurate enough to be used as training data for the ConvLSTM-based SWH prediction model. Model A was preliminarily established by only using the SWH from WWIII as the training data, and 20 sensitivity experiments were carried out to investigate the influences of different parameter settings on the forecasting effect of Model A. The experimental results showed that Model A has the best forecasting effect when using three years of training data and three hourly input data. With the same parameter settings as the best prediction performance Model A, Model B and C were also established by using more different training data. Model B used the wind shear velocity and SWH as training and input data. When making a 24-h SWH forecast, compared with Model A, the root mean square error (RMSE) of Model B is decreased by 17.6%, the correlation coefficient (CC) is increased by 2.90%, and the mean absolute percentage error (MAPE) is reduced by 12.2%. Model C used the SWH, wind shear velocity, wind and wave direction as training and input data. When making a 24-h SWH forecast, compared with Model A, the RMSE of Model C decreased by 19.0%, the CC increased by 2.65%, and the MAPE decreased by 14.8%. As the performance of the ConvLSTM-based prediction model mainly rely on the SWH training data. All the ConvLSTM-based prediction models show a greater RMSE in the nearshore area than that in the deep area of SCS and also show a greater RMSE during the period of typhoon transit than that without typhoon. Considering the wind shear velocity, wind, and wave direction also used as training data will improve the performance of SWH prediction.

In this paper, a 2-stage cascaded deep learning framework, Port Wave Prediction Network (PWPNet), is proposed for real-time prediction of significant wave height (SWH) distribution in a port. The PWP-out model of the first stage, predicting port-entrance wave parameters, utilizes three branches, the first branch using a Long Short Term Memory (LSTM) module to learn the temporal dependencies of time sequences of port-entrance wave parameters, the second branch using Wave and Wind field Feature Extraction (WWFE) modules, composed of a residual network with spatial and channel attention, to capture spatiotemporal characteristics of outside-port 2D wave and wind field data, the third branch using multi-scale time encoding to capture the periodic characteristics of waves and wind. The PWP-in model of the second stage, estimating the in-port SWH distribution, uses port-entrance wave parameters based on a customized Artificial Neural Network (ANN) and takes PWP-out’s output as its input. A comparison of the performance of PWP-out and mainstream machine learning models including LSTM, GRU, BPNN, SVR, ELM, and RF at Hambantota Port shows that PWP-out outperforms all other models regarding medium-term (25–48 h), med–long-term (49–72 h), and long-term (73–96 h) predictions, and ablation experiments proved the effectiveness of the three branches. Furthermore, the performance comparison of our PWPNet and other 2-stage models of LSTM, GRU, BPNN, SVR, ELM, and RF cascaded with PWP-in shows that PWPNet outperforms those cascaded models for medium-term to long-term predictions of SWH distribution in a port.

Significant wave height (SWH) prediction plays an important role in marine engineering areas such as fishery, exploration, power generation, and ocean transportation. For long-term forecasting of a specific location, classical numerical model wave height forecasting methods often require detailed climatic data and incur considerable calculation costs, which are often impractical in emergencies. In addition, how to capture and use the dynamic correlation between multiple variables is also a major research challenge for multivariate SWH prediction. To explore a new method for predicting SWH, this paper proposes a deep neural network model for multivariate time series SWH prediction—namely, CLTS-Net. In this study, the sea surface wind and wave height in the ERA5 dataset of the relevant points P1, P2, and P3 from 2011 to 2018 were used as input information to train the model and evaluate the model’s SWH prediction performance. The results show that the correlation coefficients (R) of CLTS-Net are 0.99 and 0.99, respectively, in the 24 h and 48 h SWH forecasts at point P1 along the coast. Compared with the current mainstream artificial intelligence-based SWH solutions, it is much higher than ANN (0.79, 0.70), RNN (0.82, 0.83), LSTM (0.93, 0.91), and Bi-LSTM (0.95, 0.94). Point P3 is located in the deep sea. In the 24 h and 48 h SWH forecasts, the R of CLTS-Net is 0.97 and 0.98, respectively, which are much higher than ANN (0.71, 0.72), RNN (0.85, 0.78), LSTM (0.85, 0.78), and Bi-LSTM (0.93, 0.93). Especially in the 72 h SWH forecast, when other methods have too large errors and have lost their practical application value, the R of CLTS-Net at P1, P2, and P3 can still reach 0.81, 0.71, and 0.98. The results also show that CLTS-Net can capture the short-term and long-term dependencies of data, so as to accurately predict long-term SWH, and has wide applicability in different sea areas.

Explicit wave models and expensive sensor equipment capable of predicting and measuring wave parameters often carry a prohibitive computational and financial expense. To counter this, this paper proposes an alternative method for nowcasting coastal zone significant wave heights through the joint use of meteorological and structural data in the training of supervised machine learning models. In testing the hypothesis that structural data can improve model classification, artificial neural network and decision tree models were developed, trained and tested on field data recorded on a coastal jetty located in the southern coasts of Cyprus. A comprehensive investigation of the different models yields that the joint use of meteorological and structural features can improve classification performance, regardless of the network choice. It is also demonstrated that redundancy of training parameters could inject unwanted overfitting, reducing model generalization. To address this, a method for quantifying feature importance has been proposed by exploiting the nature of decision tree algorithms and the Gini impurity index, reaffirming that structural features do indeed benefit model classification. These results highlight the potential of tapping into the untapped pool of structural data for significant wave height prediction, paving the way for new research to be undertaken in this direction.

This paper proposes an alternative method for nowcasting significant wave height (Hs) through the development of hierarchical machine learning classification models. In testing the hypothesis that hierarchical classification can improve Hs prediction, flat and hierarchical classifiers were developed and tested on field-data recorded on a coastal jetty located in the southern coasts of Cyprus. A comprehensive investigation of the performance of flat over hierarchical classification models yields that the proposed method provides greater flexibility throughout the model development stages. This flexibility is attributed to the manipulation of data before training, optimization of classifier's hyperparameters during training, and the curtailment of features post-training at each level of the hierarchy. It is demonstrated that, the hierarchical approach resulted in better classification performance across a plethora of performance metrics established for a comprehensive comparison. It is also shown that the increased performance of the proposed approach comes at the expense of complexity arising from performing computationally expensive operations and the requirement for development of multiple local classifiers. Still, the increased classification performance of the hierarchical approach highlights the potential of this original method and the requirement for a rigid framework to be constructed for the development of hierarchical models for Hs prediction.