Predicting the flow velocity of tidal energy generators is a core component of tidal energy development, power assessment, unit scheduling, and operational optimization. To improve the accuracy and real-time performance of tidal velocity prediction, this paper proposes a prediction method that integrates statistical modeling and deep learning mechanisms. This method constructs an integrated model framework, VMD-AL-BRNN, specifically to model the nonlinear, multi-scale, and sudden change characteristics of tidal data. First, variational mode decomposition (VMD) is used to decompose the original tidal velocity series and extract multi-frequency characteristic components to improve data stationarity. Second, a bidirectional long short-term memory network (Bi-LSTM) is used to model the flow velocity in both the forward and reverse time directions, comprehensively capturing the temporal dependence of the velocity evolution. Furthermore, a self-attention mechanism is introduced to effectively focus on key time slices, such as high and low tide transitions, improving the model’s ability to identify local changes. Experiments based on 177 days of measured data from the Zhoushan waters demonstrate that the proposed VMD-AL-BRNN model exhibits superior prediction accuracy, stability, and generalization capabilities in a nonlinear tidal environment. In terms of quantitative evaluation metrics, the mean square error (MSE) was 0.0068 and the coefficient of determination (R2) was 0.96. Furthermore, the model has a compact structure and high training efficiency, making it suitable for deployment on edge computing platforms in tidal power generation systems, demonstrating promising prospects for industrial application.