Soft sensor-driven spatiotemporal-periodic synergistic predictive control for blast furnace gas flow

The blast furnace ironmaking process exhibits periodic behavior, time-varying delays, and complex spatiotemporal coupling, making it difficult to achieve real-time monitoring of gas flow distribution. In response to these challenges, this paper proposes a soft sensor-driven proximal policy optimization (PPO) framework with spatiotemporal periodic modeling and dynamic memory (SPDM-PPO) for synergistic predictive control. Firstly, to overcome the modeling inaccuracies caused by dynamic coupling and uncertain time delays, a dynamic time-delay optimization method is developed by embedding spatial regularization into mutual information, eliminating the hysteresis effects. Subsequently, a dual-encoding Transformer network is designed, which incorporates both absolute and periodic positional encodings to capture spatiotemporal periodic patterns and global dynamics. Then, considering the issues of information redundancy and memory obsolescence in periodic state representation, a dynamic periodic state memory (DCSM) mechanism is proposed by aggregating dual-threshold memory optimization and attention-weighted. Furthermore, to achieve dynamic closed-loop predictive control of gas flow distribution, a cooperative dual-optimizer-trained PPO strategy and the DCSM are embedded, along with a long short-term memory (LSTM) encoder-decoder. Finally, extensive experiments conducted on real-world BF industrial data robustly validate the effectiveness and superiority of the proposed framework.