Bias Correction of Operational Storm Surge Forecasts Using Neural Networks

TL;DR

This paper demonstrates that neural network-based residual learning can significantly improve operational storm surge forecasts by reducing errors, especially on short timescales, with minimal operational adjustments needed.

Contribution

It introduces a neural network residual correction method for storm surge forecasts that is easy to deploy and enhances forecast accuracy without modifying existing models.

Findings

Root Mean Square Error reduced by 36% at 1-hour lead time

Error reduction of 9% at 24-hour lead time

Method is fast, operationally feasible, and enhances forecast reliability

Abstract

Storm surges can give rise to extreme floods in coastal areas. The Norwegian Meteorological Institute produces 120-hour regional operational storm surge forecasts along the coast of Norway based on the Regional Ocean Modeling System (ROMS), using a model setup called Nordic4-SS. Despite advances in the development of models and computational capabilities, forecast errors remain large enough to impact response measures and issued alerts, in particular, during the strongest events. Reducing these errors will positively impact the efficiency of the warning systems while minimizing efforts and resources spent on mitigation. Here, we investigate how forecasts can be improved with residual learning, i.e., training data-driven models to predict the residuals in forecasts from Nordic4-SS. A simple error mapping technique and a more sophisticated Neural Network (NN) method are tested. Using the NN residual correction method, the Root Mean Square Error in the Oslo Fjord is reduced by 36% for lead times of one hour and 9% for 24 hours. Therefore, the residual NN method is a promising direction for correcting storm surge forecasts, especially on short timescales. Moreover, it is well adapted to being deployed operationally, as i) the correction is applied on top of the existing model and requires no changes to it, ii) all predictors used for NN inference are already available operationally, iii) prediction by the NNs is very fast, typically a few seconds per station, and iv) the NN correction can be provided to a human expert who may inspect it, compare it with the model output, and see how much correction is brought by the NN, allowing to capitalize on human expertise as a quality validation of the NN output. While no changes to the hydrodynamic model are necessary to calibrate the neural networks, they are specific to a given model and must be recalibrated when the numerical models are updated.