This study explores storm floods in the Yangtze Estuary to investigate how extreme sea levels and storm surges change in the context of global warming. Previous studies focused on the long-term variations in amplitude or frequency of storm surges, with limited research conducted on the timing of extreme storm surge events. Based on the methods of non-stationary extreme value theory, we explored the last 33-year tidal levels at Xuliujing Station and found that the annual extreme water level has exhibited a slight downward trend, which is directly attributed to the decrease in mean sea level resulting from reduced upstream river flow. The storm surge season of the Yangtze Estuary experienced a significant lag in the period 2005–2018, which is not restricted to the Yangtze Estuary but is rather a large-scale climate characteristic of a broad oceanic region. The reason for this shift is the sustained increase in the intensity of the western Pacific subtropical high in the last 15 years, leading to the appearance of low-pressure channels in the East China Sea in September and October and thus causing more typhoons to enter the East China Sea during the later period of the storm surge season.

Accurate reservoir outflow prediction is crucial for water resource management. Although seasonal decomposition has shown good performance in time series prediction, its application in reservoir outflow prediction remains limited. This study proposes a decomposition–prediction–reconstruction framework for reservoir outflow prediction, integrating the STL (Seasonal and Trend decomposition using LOESS) method with the Informer model. The framework comprises three steps: first, decomposing complex reservoir outflow time series into trend, seasonal, and residual components using STL; then applying the Informer model to predict each component separately; and finally, reconstructing the final prediction by summing all components. Using 64-year data (1958–2021) from the Lianghekou Reservoir in the Yalong River (China) as a case study, STL-Informer was compared with five baseline methods, including extreme gradient boosting and long short-term memory (both with and without STL integration). Results demonstrate that STL decomposition significantly improves all baseline models’ performance, with enhancement ratios ranging from 27.35 to 88.04% across different metrics. STL-Informer achieves the best performance (NSE = 0.8650), showing 48.21% improvement in NSE, 43.05% reduction in RMSE, and 45.09% reduction in MAE compared with the baseline Informer. This study provides an effective technical approach for reservoir outflow prediction with significant value for reservoir operation management.