CIP: TurboSWAN to assess confidence intervals for SWAN forecasts : surrogate North Sea wave model to apply wind ensembles in wave forecasts

Within the Rijkswaterstaat Operational Systems (RWsOS), SWAN wave models are used to provide wave forecasts for both the Dutch coast and the larger lakes. On average these forecasts are quite good but the scatter in wave height can be rather large, especially for the low frequency wave height HE10. There is a need for confidence intervals to the wave forecasts. Therefore we have tried to set up TurboSWAN, a fast surrogate model trained on SWAN in- and output, which should be able to make wave simulations on the entire SWAN-North Sea domain for 50 ECMWF wind ensemble members within minutes. The result, being the spread in wave parameters Hm0, (and ultimately Tm-1,0 and HE10) will be applied on the outcome of the one SWAN run which used the control wind field as input. Initially, the dataset for training and testing (i.e. validation) came from SWAN-North Sea runs with input from the TIGGE project. Various experiments were done to improve TurboSWAN but the error statistics did not reduce significantly. Aspects considered in these experiments were time lag of input fields, focussing on part of the domain, definition of the loss function to be minimised, data augmentation, architecture (drop out layers, up- and downsampling of data, etc). The hypothesis that the training set was too uniform seemed to be false: changing to a more varied dataset (18 selected years out of the 45 year ERA5 dataset) did not result in the desired improvement. TurboSWAN represents large-scale wave height features quite well. However, with typical RMSE values of approximately 0.6 m in significant wave height relative to the SWAN results for eight selected locations, it is not yet good enough for operational implementation. Apart from this rather large RMSE, there is no full confidence in the present set up of TurboSWAN as some experiments with a physically sound basis, did not lead to large improvements in the performance. SWAN has a typical RMSE of 0.3 m for significant wave height relative to observations (Deltares, 2023). There are various criteria to judge the experiments, i.e. RMSE, bias, for specific locations or for the entire domain and it is hard to define an overall best score. When using different training sets, the statistical scores like RMSE are not necessarily directly comparable. TurboSWAN is very fast: 25ms to compute one time step. For eight 6-hr-time steps covering 48 hours, and fifty ensembles, it would take about 10 seconds on GPU (and approximately a factor 40 more on CPU) to compute the wave field. These time-steps can even be executed in parallel, contrary to SWAN. The amount of publications on surrogate wave modelling is scarce, especially for spatial applications in a regional shelf sea. Given the convergence of improvements in the current experiments, we advise to decompose the experiments into smaller components and tests and use a modular step-by-step approach. For example, this could be combined with a timeseries-CNN (convolutional neural network) for the core stations allowing for fast experiments and transferring lessons learnt to the spatial CNN.