ML-Based Storm Surge Prediction

This repository provides the code and synthetic training data to support the reproducibility of machine learning experiments presented in the paper "Combining Synthetic and Observed Data to Improve Machine Learning Model Performance for Severe Storm Surge Forecasting along the Southeastern Coast of China. (in submission)" Dataset:The synthetic dataset was generated using the FVCOM (Finite Volume Community Ocean Model) hydrodynamic model and includes: 53 historical typhoon simulations (1959–2017) 27 perturbed experiments systematically varying typhoon parameters (central pressure, maximum wind speed, radius of maximum wind, and track position) 25 extreme scenarios representing hypothetical severe storm surge events Input features comprise typhoon center coordinates, minimum central pressure, maximum sustained wind speed, translation speed and direction, and distance to the target station. The prediction target is hourly storm surge at six tide gauge stations along China's southeastern coast. ⚠️ Observational tide gauge data referenced in the paper are not included due to data sharing restrictions under Chinese surveying and mapping regulations. The repository documents the data structure to enable users to integrate their own observational datasets. Code:The codebase includes implementations of four ML architectures—Multi-Input Multi-Output ANN, standard LSTM, CNN-LSTM, and Encoder-Decoder LSTM—together with utilities for data preprocessing, train/test splitting, model training and evaluation, and result visualization.