Smart contracts have been developing rapidly in various industries, and their security problems have become increasingly severe, posing significant economic threats. Consequently, the identification of vulnerability in smart contracts is essential for the governance of blockchain. Deep learning has achieved remarkable strides in bolstering software security, and deep learning and smart contract vulnerability detection have become a new research focus. Existing smart contract vulnerability detection methods have difficulty eliminating the influence of redundant information, leading to insufficient special extraction, which leads to missed and false alarms. To bolster the security of smart contracts, we introduce a novel technique that leverages the temporal convolutional network (TCN), specifically the HETCN model. This model refines the Abstract Syntax Tree (AST) to mitigate the impact of superfluous data, thereby enhancing feature extraction. Moreover, the TCN integrates the squeeze-and-excitation network mechanism to distinguish the weights of the data and overcome long-range dependency. We perform reentrancy vulnerability detection on the widely used public dataset SmartBugs Wild. The results indicate that our HETCN method detects reentrancy vulnerability with a higher accuracy of 95.81%. Compared with other detection methods, HETCN can better identify reentrancy vulnerability.