A Lightweight Field-to-Site Coupled Framework for 15-Day Sea Surface Temperature Forecasting in Marine Ranching Areas: A Case Study in the Northern Yellow Sea

Sea surface temperature (SST) anomalies represent a critical threat to the operational safety and productivity of marine ranching systems. Taking a typical marine ranching area in the Northern Yellow Sea as the study area, this study developed a lightweight two-stage field-to-site forecasting framework. In the first stage, a Convolutional Long Short-Term Memory network (ConvLSTM) was employed to generate 1–5 days regional SST forecasts. Through experiments involving 24 input configurations, the combination of Optimum Interpolation Sea Surface Temperature (OISST), seasonal and trend components, and ERA5 meteorological variables was identified as an optimal configuration, providing spatial-evolution constraints for the target location. In the second stage, the 5-day target-site forecasts were concatenated with OISST to construct a 60-day sequence, which was then used to drive a lightweight Gated Recurrent Unit (GRU) for extended forecasts from days 6 to 15. This strategy reformulates the forecasting task into a spatially constrained 10-day extension forecast, thereby suppressing long-lead error accumulation. The results showed that, in the regional forecasting stage, an RMSE of 0.83 °C and MAE of 0.63 °C were achieved on Day 5, while in the extended-forecasting stage, an RMSE of 0.89 °C and MAE of 0.69 °C were achieved on Day 15. Compared with single-stage ConvLSTM and GRU models performing direct 15-day forecasting, the field-to-site strategy reduced RMSE by approximately 25.21% and 9.18%, respectively, and reduced MAE by approximately 24.18% and 8.00%, respectively. Independent validation against in situ buoy observations further showed that the proposed framework introduced only limited additional errors comparable to the inherent discrepancy between OISST and buoy observations. Additional tests under representative marine heatwave (MHW) conditions showed that the framework retained useful forecasting skill under anomalous warming conditions. Furthermore, its extended application experiments at five additional marine ranching sites in the Northern Yellow Sea produced consistent forecasting performance, with mean RMSE and MAE ranging from 0.71 °C to 0.75 °C and from 0.53 °C to 0.56 °C, respectively. The proposed method can therefore provide a risk-warning window of at least two weeks for marine ranching management and support timely operational decisions for temperature-related risk mitigation and refined aquaculture management.