Fast prediction of oceanic nearshore circulation (FPOCH) system with multi-model fusion strategies

• FPOCH digital twin system merges numerical & data-driven models for oceanic forecast. • Multi-model fusion outperforms data-driven models in matching oceanic process heterogeneity. • Algorithm using sparse tidal gauge data enables nearshore current field forecast. • EMD preprocessing enhances free surface level forecasting precision. • ConvLSTM raises current field model accuracy at ebb and flood phases. Accurate representation of nearshore flow fields is essential for effective coastal management and risk mitigation. Traditional numerical models often involve a trade-off between predictive accuracy and computational efficiency. In contrast, purely data-driven ocean models struggle to adequately capture multiscale processes, regional heterogeneity, and extreme events, limiting their applicability to operational forecasting. To address these challenges, this study proposes the Fast Prediction of Oceanic Nearshore Circulation (FPOCH) system, a digital twin framework that integrates numerical simulations (FPOCH numerical simulations, hereafter FPOCH NS), a model-nesting strategy, and machine-learning (ML) techniques. This integrated framework enables rapid, accurate, and scalable predictions spanning from the continental shelf to nearshore coastal zones. FPOCH NS employs a model-nesting strategy that integrates two physics-based models and one ML platform: FPOCH NS1, which covers the China adjacent seas; FPOCH NS2, which provides higher-resolution simulations for the South Yellow Sea (SYS); and FPOCH-ML, which targets a nearshore subdomain within the SYS. At large spatial scales, FPOCH NS outperforms global operational models in forecasting accuracy, thereby providing boundary condition and forcing data for the subsequent rapid forecasting framework implemented in FPOCH-ML. At the nearshore scale, FPOCH-ML incorporates advanced algorithms for rapid current prediction, leveraging sparse tidal observations while maintaining an update frequency consistent with tide-gauge measurements. FPOCH-ML includes two algorithmic modules: a single-point forecasting algorithm and a field forecasting algorithm. For each module, multiple ML models are trained to operate collaboratively, and a range of network architectures is systematically tested and compared. Results indicate that a hybrid ML architecture based on Empirical Mode Decomposition (EMD-ML) outperforms long short-term memory (LSTM) and convolutional LSTM (ConvLSTM) models in tidal level forecasting for both single-point and field applications. By contrast, LSTM and ConvLSTM demonstrate superior performance relative to EMD-ML in current forecasting across both application scenarios. The framework delivers high predictive accuracy after optimization: EMD-ML achieves an RMSE of 0.03 m for tidal level forecasting; LSTM an RMSE of 0.03 m/s for single-point current velocity forecasting; and ConvLSTM an RMSE of 0.03 m/s for current velocity and 5.46° for current direction in field current forecasting. Notably, FPOCH NS yields errors approximately three times lower than those of a global model during a typhoon-induced storm surge, while FPOCH-ML operates about 20 times faster than a traditional numerical simulation. This study demonstrates that multi-model fusion, integrating physics-based and data-driven approaches, outperforms both standalone data-driven models and conventional global operational systems. FPOCH thus provides an efficient framework for ensemble forecasting across multiple spatial and temporal scales, delivering actionable insights for nearshore hydrodynamic applications, disaster early warning, and coastal engineering design.