Abstract: To address the challenges of capturing non-stationary characteristics of quality indicators and insufficient long-term dependency modeling accuracy in batch production processes, this study proposes a Bidirectional Temporal Difference Memory Network （BiTDMN）, which integrates local dynamic feature extraction and global temporal modeling capabilities. Three-dimensional batch data are reconstructed using sliding windows; key variables are selected via mutual information; non-stationary feature extraction is enhanced through forward-backward difference operations and a bidirectional recursive structure is employed to achieve collaborative modeling of historical and future information. An incremental online update mechanism is introduced to fine-tune the model every 20 hours using offline detection data, suppressing multi-step prediction bias accumulation. Validation on industrial penicillin fermentation data shows that BiDRNN achieves a mean absolute error （MAE） of 0.103, root mean square error （RMSE） of 0.131, and coefficient of determination （R²） of 0.972. Long-period prediction results demonstrate a significant reduction in fitting errors for concentration peaks, with a 60% reduction in cumulative deviation for multi-batch endpoint predictions. The online update strategy achieves near-zero error in inflection point time prediction, verifying the model’s effectiveness in non-stationary dynamic modeling and real-time correction. This method provides an efficient solution for online quality monitoring in batch processes across industries such as fine chemicals and biopharmaceuticals.