Transient voltage stability is of significant importance for the stable operation of power systems. To swiftly and accurately assess the transient voltage stability condition following a fault in the power system, an evaluation method combining attention mechanism (AM) and bidirectional long short-term memory network (BiLSTM) is proposed. Firstly, the transient input data representing the operating state of the power system are constructed by the electrical quantity of each node in the three stages of pre-fault, fault and post-fault. Then, the BiLSTM deep neural network algorithm is constructed to fully capture the time series features of transient data, and the AM is incorporated for assigning feature weights, thereby increasing focus on the most significant ones. In addition, to improve the model tendency problem caused by the inherent class imbalance of transient samples, the weighted cross entropy loss function (Wce) is used to supervise model training in the training process. Ultimately, the IEEE 39-bus system is considered as a case for validating the precision and efficiency of the proposed approach.