The fluctuation of propulsion load is a critical factor affecting the stability of the power system and the optimal energy distribution in hydrogen-powered ships. To address the challenges posed by highly fluctuating operating conditions in hydrogen-powered ships, this paper proposes a short-term prediction model for propulsion load. First, an improved temporal convolutional network (ITCN) is employed to extract sequential features from the input data. Then, a bidirectional long short-term memory (BiLSTM) network is utilized to capture both the long-term dependencies and bidirectional correlations of the propulsion load. Finally, an attention mechanism is introduced to enhance the model’s ability to focus on critical information, thereby improving prediction accuracy. Simulation results demonstrate that the proposed model achieves significantly higher prediction accuracy across three different time scales compared to other benchmark models.