Network traffic represents the volume of data sent and received during online website visits. Anomalies in network traffic indicate unusual variations in traffic, which are crucial to detecting timely and accurately in complex computer network systems to ensure efficient operation. Existing methods for anomaly detection in network traffic have rarely focused on effectively handling time-series data in the temporal dimension. Addressing the limitations of traditional Temporal Convolutional Networks in capturing local and significant features of time series, this paper proposes a network traffic anomaly detection method based on CAT-BiLSTM. CNN-Attention(CA) extract local features from sequences, while TCN learns abstract and high-level sequence patterns to capture long-term dependencies and hierarchical features. Bidirectional Long Short-Term Memory Networks (BiLSTM) capture long-term dependencies from both directions, incorporating attention mechanisms (Attention) in both TCN and BiLSTM modules. The CAT-BiLSTM model performs bidirectional temporal modeling of sequence features, capturing long-term dependencies and temporal patterns, thereby enhancing the capability to detect anomalous behaviors in network traffic. Experimental results indicate that the proposed method significantly improves anomaly detection performance compared to traditional machine learning approaches, showing more stable and accurate handling of time-series data.