[Paper Review] Storm Surge Modeling in the AI ERA: Using LSTM-based Machine Learning for Enhancing Forecasting Accuracy

Stefanos Giaremis, Noujoud Nader+4 more|arXiv (Cornell University)|Mar 7, 2024

Meteorological Phenomena and Simulations4 citations

TL;DR

This paper proposes an LSTM-based deep learning model to predict and correct systemic biases in physics-based storm surge forecasts using historical gauge observations. Trained on 61 past hurricanes, the model improves forecast accuracy for Hurricane Ian (2022) by correcting simulated water levels, achieving R² > 0.9 for short windows and maintaining R² ~0.7–0.8 for 18-hour predictions, with minimal data requirements—only six diverse storms needed for comparable performance to full-dataset training.

ABSTRACT

Physics simulation results of natural processes usually do not fully capture the real world. This is caused for instance by limits in what physical processes are simulated and to what accuracy. In this work we propose and analyze the use of an LSTM-based deep learning network machine learning (ML) architecture for capturing and predicting the behavior of the systemic error for storm surge forecast models with respect to real-world water height observations from gauge stations during hurricane events. The overall goal of this work is to predict the systemic error of the physics model and use it to improve the accuracy of the simulation results post factum. We trained our proposed ML model on a dataset of 61 historical storms in the coastal regions of the U.S. and we tested its performance in bias correcting modeled water level data predictions from hurricane Ian (2022). We show that our model can consistently improve the forecasting accuracy for hurricane Ian -- unknown to the ML model -- at all gauge station coordinates used for the initial data. Moreover, by examining the impact of using different subsets of the initial training dataset, containing a number of relatively similar or different hurricanes in terms of hurricane track, we found that we can obtain similar quality of bias correction by only using a subset of six hurricanes. This is an important result that implies the possibility to apply a pre-trained ML model to real-time hurricane forecasting results with the goal of bias correcting and improving the produced simulation accuracy. The presented work is an important first step in creating a bias correction system for real-time storm surge forecasting applicable to the full simulation area. It also presents a highly transferable and operationally applicable methodology for improving the accuracy in a wide range of physics simulation scenarios beyond storm surge forecasting.

Motivation & Objective

• To develop a machine learning framework that predicts and corrects systemic errors in physics-based storm surge models.
• To enhance forecasting accuracy post facto by applying learned bias corrections to simulation outputs.
• To evaluate whether a small, representative subset of historical storms can yield performance comparable to full-dataset training.
• To enable operational deployment of bias correction in real-time storm surge forecasting systems.
• To demonstrate transferability of the method to other physics simulation scenarios beyond storm surge modeling.

Proposed method

• The model uses Long-Short Term Memory (LSTM) networks to learn temporal patterns in offset time series between observed and modeled water levels.
• Offsets were computed as the difference between observed water levels from NOAA/USGS gauge stations and ADCIRC-simulated levels.
• A sliding window approach was applied to convert offset sequences into supervised learning training samples.
• Data were standardized and split into training and validation sets prior to model training.
• The model was trained on 61 historical hurricanes and tested on Hurricane Ian (2022), which was excluded from training.
• Performance was evaluated by comparing corrected predictions ('Modeled+ML') against raw model outputs ('Modeled') using R² and visual analysis across gauge stations.

Experimental results

Research questions

• RQ1 Can an LSTM-based model effectively predict systemic errors in storm surge forecasts using historical offset data?
• RQ2 How does model performance vary when trained on a reduced subset of hurricanes, particularly those with diverse tracks relative to the target storm?
• RQ3 To what extent can a small, representative subset of storms (e.g., six) achieve performance comparable to full-dataset training?
• RQ4 Does the model maintain robustness across different prediction horizons, including longer windows up to 18 hours?
• RQ5 Can the model be operationally deployed in real-time forecasting with minimal retraining and low computational overhead?

Key findings

• The LSTM model achieved R² > 0.9 for prediction windows up to 6 hours and R² ~0.7–0.8 for 18-hour prediction windows when correcting ADCIRC-simulated water levels for Hurricane Ian.
• Applying the ML-predicted offsets to the model output ('Modeled+ML') significantly improved forecast accuracy at the vast majority of gauge stations compared to the raw model.
• A training subset of just six hurricanes with diverse tracks relative to Hurricane Ian produced bias correction performance statistically indistinguishable from the full 61-storm dataset for prediction windows of 9 hours and longer.
• The Wilcoxon signed-rank test confirmed no statistically significant difference in R² performance between the full-dataset and six-storm subset models across multiple input window sizes.
• Training time scaled linearly with both the number of data points and the input window size, indicating predictable and manageable computational costs.
• The model demonstrated potential for operational use, as it relies solely on historical offset data and does not require real-time physics model re-execution.