Driven by the “dual carbon” strategy and the high-quality development of the Hainan Free Trade Port, accurate and interpretable forecasting of port carbon emissions is urgently needed to overcome the limitations of traditional models. To tackle the challenges of multi-factor modeling, the traditional AutoRegressive Integrated Moving Average (ARIMA) model is first used as a baseline forecast. A hybrid neural network framework is then proposed, combining Convolutional Neural Networks (CNN) for feature extraction, Long Short-Term Memory (LSTM) networks for temporal relationships, and eXtreme Gradient Boosting (XGBoost) for non-linear error correction. To enhance interpretability, the SHapley Additive exPlanations (SHAP) method quantifies the impacts of variables like port operations and energy structure. These insights are integrated into a System Dynamics (SD) feedback model to construct a causal loop linking emissions, operations, policy, and technology. An empirical study on Yangpu Port in Hainan confirms the proposed approach achieves high accuracy and strong generalizability, supporting intelligent forecasting and policy intervention in complex port emission systems.