Machine learning based accurate storm surge peak and timing forecast in Pearl River Estuary

With climate warming, storm surge increasingly threatens coastal cities, underscoring the necessity for accurate and timely forecasts of peak surge height and timing. Among widely used neural network-based approaches in storm surge forecasting, the Gated Recurrent Unit (GRU), while strong in sequential prediction, tends to underestimate peak magnitudes and exhibits timing errors at extended lead times. To address the issue, this study developed a single-station GRU prediction model using storm surge data from 42 tropical cyclones (2000-2023) at nine stations in the Pearl River Estuary. To reduce peak underestimation, an iterative forecasting scheme was designed, incorporating tropical cyclone (TC) parameters selected via Accumulated Local Effects (ALE) and correlation coefficient. To mitigate peak timing errors, a Support Vector Regression (SVR) model based on TC peak intensity elements was proposed to predict and correct peak timing errors. Additionally, multi-station joint forecasting was explored by embedding spatial information between stations and TC tracks. Results based on leave−one−out cross−validation (LOOCV) indicated that integrating TC location, distance, and 34−kt wind radii in the key quadrants yielded a 27% reduction in Peak Error (PE) at 12–18 h and a 13% reduction in RMSE over 1–18 h for the 35 test cases. After applying the SVR−based phase correction, the forecast for all cases achieved an 55% reduction in Timing Error (TE) and a 2.81% reduction in RMSE relative to the uncorrected forecast. Compared to numerical model outputs, the RMSE of water level sequence during Typhoon Hato was decreased by 0.22 m. Multi-station joint forecast results showed that embedding such spatial information not only enhances forecast skill across multiple gauges but also provides benefits for stations with sparse historical data. This study offers a promising method that targets both the magnitude and timing of storm surge peaks, thereby improving the accuracy and reliability of storm surge forecasting.