Abstract
Western boundary currents offer a persistent source of marine renewable energy, yet their deployment remains limited by the lack of accurate, site-scale forecasts that can be translated into electricity. Transformer-based artificial intelligence has transformed prediction, but its translation into site-specific ocean-current power forecasting for real-world energy applications remains strikingly limited. Here we develop an observation-constrained hybrid deep-learning (HyDL) framework that advances Kuroshio prediction toward grid-relevant, ocean-current energy applications. This study delivers four key contributions. First, we develop a hybrid Transformer-based AI framework that improves site-scale Kuroshio current prediction across multiple spatiotemporal scales. Second, by synergistically integrating global self-attention, statistical correction, and observation-driven real-time learning, HyDL achieves an average ∼ 30% reduction in RMSE relative to a state-of-the-art operational system. Third, predicted currents are translated into turbine-scale power output under realistic operational constraints, and providing a direct pathway to reduce generation uncertainty and system-level cost. Finally, this work provides a scalable blueprint that integrates Kuroshio power generation, Transformer-based AI forecasting, and Taiwan’s energy management system, accelerating the operational deployment of ocean-current energy in Taiwan and across major western boundary current regions worldwide.