Improving significant wave height prediction via temporal data imputation

Abstract

Accurate prediction of significant wave height (SWH) is crucial for a wide range of marine and coastal applications. However, achieving an accurate data-driven prediction of SWH requires effective multivariate time series modeling. Furthermore, missing values appear frequently in the raw data and influence the accuracy of the prediction. In this study, we propose a novel diffusion-based approach for continuous-time modeling and temporal imputation of multivariate time series. By learning the temporal correlations and interdependencies among variables in the buoy’s data, the imputation of missing data is conducted to enhance the SWH prediction. Experiments are performed using buoy data from the National Data Buoy Center of USA to validate the effectiveness of temporal imputation and the use of multivariate data. The experimental results, compared with baseline methods and univariate predictions, highlight the advantage of Conditional Score-Based Diffusion Models (CSDI) in capturing temporal correlations and its effectiveness in improving short-term predictions of SWH. CSDI improves imputation by 7%–30% over existing imputation methods on popular performance metrics. Compared to univariate data, the better SWH prediction results on multivariate data confirm that temporal data imputation is beneficial for prediction.