In the modern industrial process, soft sensor modeling is crucial in solving the problem of unmeasurable key quality indicators and achieving real-time dynamic monitoring. However, traditional data-driven methods struggle to capture the complex spatial and temporal dynamics in industrial time-series data. This study proposed a hybrid deep learning framework that integrates temporal convolutional network (TCN) for robust spatial feature extraction with long short-term memory (LSTM) networks for effective temporal dependency modeling. By incorporating batch and control input embeddings, the model adeptly handles inter-batch variability and diverse operational conditions. Evaluated on sulfur recovery and penicillin fermentation datasets, the proposed TCN-LSTM framework demonstrates superior predictive accuracy and robustness compared to baseline models, including standalone TCN, LSTM, and other neural architectures.