Short-term high-volatility power load forecasting in smart port energy systems using FeatureGating-BiLSTM enhanced by DualAttention mechanisms

Modern seaports are transforming into complex integrated energy hubs, deploying renewable energy, storage, and advanced electrification to achieve carbon neutrality. Accurate multi-energy load forecasting in ports is crucial for optimizing energy scheduling, improving efficiency, and reducing operating costs. However, conventional models often struggle with the highly volatile, coupled load patterns in port energy systems. To address this gap, a machine learning based model for forecasting port energy demand has been developed. The architecture integrates a dynamic FeatureGating mechanism that adaptively emphasizes informative predictors while attenuating irrelevant noise in real time. The bidirectional sequence encoder captures temporal dependencies in load time-series, processes the FeatureGating outputs, and furnishes rich temporal patterns for the subsequent attention modules. This study introduces the first port-oriented forecasting framework that integrates a DualAttention scheme with an output-level noise re-parameterization layer, while concurrently fusing fine-grained meteorological variables and real-time grid-side data within a unified DeepLearning pipeline. Specifically, the model pioneers an original DualAttention mechanism, its LocalAttention module precisely captures fine-grained, short-term fluctuations, whereas its FeatureAttention module uncovers the complex cross-feature interactions essential for port operations. To quantify forecast uncertainty and yield actionable probabilistic estimates, an output-noise reparameterization strategy is integrated into the prediction head. In a comprehensive real-world case study based on operational data, the method outperformed multiple baseline forecasters, reducing the Mean Absolute Percentage Error from 13.75 % to approximately 11.12 %. These findings underscore the methodological contribution and practical utility of the approach.