This paper proposes a deep learning-based surrogate model to accelerate the transient simulation of on-board bidirectional DC-DC converters in electric vehicles (EVs). Traditional SPICE-based transient simulation of power electronic converters is computationally expensive, particularly when extensive parameter sweeps and design iterations are required during the development of EV power distribution systems. To address this limitation, a long short-term memory (LSTM) neural network is trained to predict the transient output voltage and inductor current waveforms of a bidirectional Buck-Boost converter. A dataset of 8,000 simulation cases is generated using LTspice, covering practical ranges of circuit parameters including inductance (20-150 uH), capacitance (20-150 uF), load resistance (1-15 ohm), switching frequency (50-150 kHz), and input voltage (24-48 V). The trained surrogate model achieves a root mean square error (RMSE) of approximately 1.1% for output voltage prediction and 2.3% for inductor current prediction on the test set, while providing a computational speedup of about 70 times for single-sample inference and over 200 times for batch processing compared to conventional SPICE simulation. The model shows reasonable generalization within the training parameter range, and the incorporation of physics-informed regularization helps reduce physically inconsistent predictions. The results indicate that the proposed approach can serve as a useful auxiliary tool for rapid design space exploration in the early stages of EV power electronics development.