<p>In vehicle-to-vehicle (V2V) communication scenarios, the high-speed movement of vehicles causes channels to exhibit dynamic characteristics, including short coherence times and strong frequency-domain spatial correlations. Traditional IEEE 802.11p channel estimators, which rely on fixed channel statistics, struggle to track these time-varying channel patterns in real time, resulting in a significant increase in the bit error rate (BER) during signal demodulation at the receiver. Concurrently, existing deep learning (DL)-assisted channel estimation methods face limitations: they either suffer from degraded performance due to channel time variability and frequency selectivity or incur excessively high computational complexity. This paper introduces a novel channel estimation model, designated GAT-TCN-TA, which integrates Graph Attention Networks (GAT) and Temporal Convolutional Networks (TCN) in parallel to extract spatiotemporal features, while incorporating temporal averaging (TA) to further suppress noise. The model employs GAT to analyze and capture the non-uniform spatial correlations of the channel in the frequency domain, combined with a TCN to capture the long-term dynamic dependencies of the channel in the time domain, effectively improving the accuracy and robustness of channel estimation. Experimental results demonstrate that GAT-TCN-TA significantly outperforms existing methods in both high-speed and low-speed scenarios. For example, in a high-speed scenario at 35 dB, its BER is approximately one order of magnitude lower than that of the second-best model, and the normalized mean square error (NMSE) is reduced by more than 50%.</p>