Forecasting sea level maxima using Machine learning with explainability and extreme value analysis

• Bayesian-optimized ML (RF, MLP, XGB) and DL (CNN-GRU, CNN-LSTM) models developed for daily Baltic Sea SLM forecasting. • Neural networks outperformed tree-based models; CNN-GRU showed best generalization (7–15 cm RMSE). • Models captured storm-driven extremes across multiple alarm thresholds using precision, recall, and F1-score. • Explainability analysis revealed regional drivers: prefilling dominates, with pressure/wind in the east and BSI in the west. • Rare extremes >150 cm were underestimated; EVT shows 5–7 yr winter return periods of these extremes in the NE Baltic. Accurately forecasting sea level maxima (SLM) and extremes, is vital for maritime management, engineering, and navigation. Most Machine learning (ML) models focus on moderate surges and often underestimate extremes. We use a two-fold forecasting framework: ML/DL for short-term daily SLM forecasting and extreme value theory (EVT) for long-term extremes (<100 years). ML models include Random Forest, Extreme gradient boosting, and Multilayer perceptron, CNN-LSTM and CNN-GRU. Long-term extremes are analysed via EVT using a block maximum method. The Baltic Sea, a semi-enclosed micro-tidal basin prone to extremes, serves as case study. The analysis used six tide gauge stations (Narva, Ristna, Oulu, Kungsholmsfort, Władysławowo, and Greifswald). Key features—wind speed, surface air pressure, Baltic Sea Index (BSI), and significant wave height (SWH), were selected using a mutual information and models’ hyperparameters tuned using Bayesian optimization. Neural networks models, specifically the CNN-GRU and MLP, performed best (RMSE 7–15 cm) with strong generalization. Most models captured storm events, but underestimated extreme peaks (>150 cm), due to the rarity in the training, incomplete meteorological representation, and missing local physical processes. CNN-GRU excelled in RMSE, recall, and F1, while MLP led in R 2 and precision. EVT analysis showed winter extremes have ∼ 5–7 years in the north-east (Narva and Oulu). Explainability analysis of CNN-GRU showed prefilling dominates SLM at all stations; BSI, pressure, and winds drive west, south, and north, while local pressure, wind, and SWH dominate in the east. The framework supports early warning and long-term risk assessment, though forecasting rare extremes remains challenging.