Filling gaps in significant wave height time series records using bidirectional gated recurrent unit and cressman analysis

Abstract

Continuous records of wave data are of great significance for marine researches and analyses. However, the measurement of wave by buoys is often interrupted for various reasons. In this paper, five machine learning methods were employed to fill in the missing significant wave height (Hs) data: random forest (RF), support vector machine (SVM), artificial neural network (ANN), gated recurrent unit (GRU) and bidirectional gated recurrent unit (BiGRU). Inputs to the model are the relevant meteorological data from one or more buoys in the vicinity of the target buoy. Then, this study adopted a data assimilation method, Cressman analysis (CA), to correct the outputs of the models. The performances of the five models before and after correction (RF-CA, SVM-CA, ANN-CA, GRU-CA, BiGRU-CA) were compared horizontally and vertically. The statistical results for the test period showed that BiGRU significantly outperforms the other four models, with lower RMSE and scatter index, and higher correlation coefficients. By CA, the errors of the five models were reduced to different degrees, and BiGRU has the most reduction. Multiple tests showed that the performance of BiGRU-CA was more stable than that of RF-CA, SVM-CA, ANN-CA and GRU-CA, and its estimation of storm waves was also accurate and reliable.