Wave condition prediction and uncertainty quantification based on SG-MCMC and deep learning model

Abstract

Deep learning methods have increasingly gained popularity in wave prediction; however, previous works predominantly focused on point estimates without quantifying the uncertainty of predictions. In certain scenarios, generating probabilistic forecasts with credible intervals is crucial for vessel navigation planning and marine infrastructure construction, etc. Therefore, this paper attempts to conduct a benchmark study on wave prediction uncertainty from both frequentist and Bayesian perspectives. A wave prediction model based on the ConvLSTM combined with various uncertainty quantification methods has been established, and a unified evaluation framework has been constructed using statistical decision theory. We utilized the fifth-generation atmospheric reanalysis dataset (ERA5) from the European Centre for Medium-Range Weather Forecasts as experimental data, employing key wave parameters (significant wave height, mean wave period, and mean wave direction) for model training and testing. Through extensive repeated experiments, we found that Bayesian methods generally provide more accurate average predictions, with the SG-MCMC method performing the best, achieving RMSE values of 0.37 m and 0.49 s for HS and TM at the 24-hour prediction, respectively, and CC values of 93.65 % and 90.65 %. On the other hand, confidence intervals obtained from frequentist methods offer broader coverage for data variations, with the MIS regression method performing the best, yielding the lowest MIS scores of only 2.92 m and 1.93 s for HS and TM at the 24-hour prediction, respectively. Indeed, this result underscores a limitation of current methods: the confidence intervals do not adequately reflect the accuracy of the predictions. In contrast, the SG-MCMC method demonstrates superior performance by excelling in both mean prediction and confidence limits.