During the high-water-cut stage of waterflood development, the pronounced nonlinearity of reservoir seepage fields and the spatiotemporal non-stationarity of injection-production dynamics present significant challenges for production forecasting. Current frequency-domain feature modeling methods are limited by their sluggish response to abrupt dynamic changes, inherent imbalance between capturing local transients and global periodicities, and susceptibility to feature coupling when processing non-stationary data. To address these issues, this paper proposes a dynamically perceptive hybrid deep-learning model named DynaTCN-Wave-BiLSTM-CA. The model incorporates a Dynamic Temporal Convolutional Network with input-adaptive dilation coefficients and channel attention to dynamically adjust receptive fields for cross-scale feature extraction. It further employs third-order db4 wavelet packet decomposition coupled with a dual time-frequency attention mechanism to achieve precise multi-scale feature decoupling. A bidirectional cross-attention fusion module based on BiLSTM is also integrated to capture spatiotemporal dynamics and delayed responses, effectively combining transient time-domain features with decoupled frequency components such as low-frequency seepage trends and high-frequency equipment noise. Validation using actual data from an offshore oilfield confirms the model's superior performance in non-stationary production sequence prediction compared to mainstream methods.