A Comparative Ablation Study of CNN-LSTM-GRU and MAformer Architectures for Operational Multi-regime Salinity Forecasting in the Outer Shannon Estuary

Accurate Sea Surface Salinity (SSS) forecasting is critical for operational management in macrotidal environments, yet the transition between local tidal forcing and long-term climatological drivers remains poorly understood in deep learning. This study presents a multi-regime evaluation and comparative ablation analysis of a CNN-LSTM-GRU (Hybrid) model versus a Multi-Head Attention Transformer (MAformer) in the Outer Shannon Estuary. Leveraging nine robust predictors categorized into Hydrodynamic, Atmospheric, and Steric/Thermal groups, an "endurance test" across horizons from 24 to 240 hours was conducted. To ensure physical consistency, a scaled lookback strategy (L = h + 1) was implemented, providing models with up to 19.4 M2 tidal cycles of historical context. Results demonstrate a distinct architectural crossover. The Hybrid model provides superior short- and medium-range stability (24h R ²  = 0.892 and 120h R ²  = 0.518), yet reaches architectural de-coherence at 240 hours, characterized by a performance collapse (R ²  = 0.206). Conversely, the MAformer exhibited superior long-term resilience, achieving lower error magnitudes (RMSE: 0.405 vs. 0.438) at the 10-day horizon. The thematic ablation reveals a scale-dependent regime shift: short-term forecasts are dominated by local velocity signals, whereas 240h stability is entirely dependent on global Steric/Mass and Thermal "Climatological Anchors". Without these anchors at 240h, both architectures experience total predictive failure; notably, the MAformer’s R ² dropped from ~ 0.216 to 0.003. Findings suggest that operational estuarine systems should adopt a hierarchical modelling approach: deploying Hybrid units for daily navigational safety and Attention-based architectures for long-term strategic planning to maintain physical-mathematical alignment across expanding temporal scales.