Thanks to the global popularity of short video apps, the video streaming service has become the most dominant traffic in mobile networks. However, the strong demands on large throughput and real-time efficiency make the video streaming service hard to maintain its high Quality of Experience (QoE), especially during high mobility events, like high-speed trains and subways. Therefore, it is crucial to analyze and predict the QoE beforehand to ensure quick adjustments. However, while existing machine/deep learning approaches analyze historical temporal patterns to forecast network conditions, they suffer from two critical limitations: (1) insufficient emphasis on temporally decaying features in recent observations, and (2) interference from pervasive video preloading mechanisms that distort real-time stall detection. To address these challenges, we propose TD-LTNet (Temporal-Decay LSTM-Transformer Network), a hybrid framework integrating sequential modeling with adaptive attention mechanisms for subway network environments. Specifically, our framework includes a Temporal-Decay Multi-Head Attention (TDMA) mechanism that effectively prioritizes time-sensitive patterns through learnable decay factors, and an Adaptive Window Attention Pooling (AWAP) module that dynamically segments input sequences and suppresses preloading-induced pseudo-features via multi-scale attention fusion. Experiments demonstrate superior prediction accuracy and real-time performance compared to conventional methods, offering an efficient solution for QoE optimization in subway network systems.