Prediction method for small-sample ocean wave data based on optimal decomposition

In marine wave prediction scenarios, where multivariate elements are relatively scarce, sample data volumes are small, and data sources are limited, a hybrid prediction model is proposed to enhance the utilization efficiency and accuracy of small-sample wave data. This model addresses challenges such as long-range dependence, extreme value fluctuations, and non-stationarity in the prediction of significant wave height time series by combining information-optimal decomposition with artificial intelligence-based prediction. By integrating an Improved Energy Valley Optimizer (IEVO) with variational mode decomposition (VMD), the model effectively extracts the feature information structure from small-sample wave data. Subsequently, a deep architecture of a Temporal Convolutional Network (TCN) enhanced with self-attention mechanisms is optimally configured based on the number of features. Finally, Bidirectional Long Short-Term Memory (BiLSTM) is employed to capture bidirectional time series correlations and short-term dynamics, resulting in more accurate predictions. The results demonstrate that the proposed model trend aligns closely with the measured values, exhibiting the smallest deviation in 96-step predictions. Compared to the existing combined model, it presents distinct advantages. In both small and large wave prediction scenarios, the MAE, MAPE, MSE, and RMSE of the proposed method improve by 25.5% to 44.6% and 8.8% to 16.9%, respectively, when compared to the corresponding optimal comparison model.