Performance and Comparison of Seq2Seq and Transformer Model Architectures for the Prediction of Water Levels from Hours to Days

Short-term water level predictions from hours to a few days are very helpful for beach managers as the information allows to prepare for beach inundation improving the safety of the public and helping to protect assets on the beach. Along the Texas coast decisions to help keep citizens safe include closing road access to local beaches. Tidal predictions are helpful in general but they do not include weather forcings and are hence are not sufficient to predict water levels during periods of strong winds, alongshore currents, large waves including such forcings generated due to an approaching storm. Using machine learning allows to combine different inputs such as past water levels, tidal predictions, past and predicted wind and wave conditions to make local predictions where a history of water levels exist. Several AI methods are compared to predict water levels with lead times from 12 hours up to 4 days. We propose the use of machine learning architecture in combination with a deep understanding of water-level physics to improve upon tidal and prior machine learning predictions. We compared different machine learning architectures, such as Seq2Seq and transformers, to achieve the best performance. This work was able to achieve a remarkable improvement compared to tidal predictions and a substantial one compared to prior machine learning architectures applied to the same coastal region.

The study area is the Gulf of Mexico, a region where many of the largest US ports are located. Several locations were selected including locations on the open coast, in embayments and ship channels, to ensure that the models were able to generalize to most of the Texas ports. The model was tested with high success for the tide gauges of Bob Hall Pier, Port Isabel, Rockport, Manchester, and North Jetty along or near the waterways of the ports of Corpus Christi, Brownsville, and Houston/Galveston. All these ports would highly benefit from more accurate operational real-time models. The goal to eventually transition these models to operation was accounted for for in the model design. For example sensors data can become unavailable, particularly during extreme events or due to maintenance or accidents hence we limited the number of inputs and length of past time series to increase the robustness of the predictions when implemented. The final models only include past wind and water level measurements and wind predictions.

This research compared the performance of multiple architectures such as Seq2Seq, and different types of transformers. From that comparison, we found that the best architecture for the described problem was achieved using Seq2Seq. Using this architecture, it was possible to maintain the Central Frequency of 15 cm (CF(15cm)) above 90% for up to 72 hours for most of the tested locations. This is a significant improvement compared to the present state-of-the-art AI models, which were only able to maintain the CF(15 cm) above 90% for 48 hours predictions or more for very few locations. Also, we found that for some locations, such as Port Isabel and Rockport, we were able to maintain a CF(15 cm) above 90% for up to 108 hours. This is a major improvement compared to the previous published work. This improvement will be highly valuable to stakeholders who will be able to benefit from these longer lead-time predictions.