Predicting time series of water levels using advanced AI models

Physics and mathematics allow us to provide accurate weather forecasts. Thanks to observations and numerical weather predictions, meteorologists can predict weather conditions which undoubtedly affect the daily choices made within various sectors of society. Models are built to provide forecasts to simulate atmospheric conditions and predict how they will evolve over time. By combining this information with hydrological data and models meteorologists can make predictions about water levels as well. However, various biases and errors are present in weather prediction models which are difficult to eliminate. We propose using machine learning as a post-processing technique to correct ocean circulation model outputs. In this research we carried out boosting algorithms together with a novel Fully Convolutional Network for the regression problem to see to what extent it is possible to successfully correct data coming from the ADCIRC model. We managed to obtain promising results which show that the model’s outputs could be successfully corrected depending on the season of the year and on number of variables which were used to train the model.