BTPNet: A Probabilistic Spatial-Temporal Aware Network for Burn-Through Point Multistep Prediction in Sintering...

Area of Science:

• Metallurgical Engineering
• Process Control
• Artificial Intelligence

Background:

• Burn-through point (BTP) is critical for sinter ore yield and quality.
• Traditional soft-sensor models struggle with the time-varying, multivariable nature of sintering.
• Existing methods face challenges in spatial-temporal feature extraction and multistep prediction error accumulation.

Purpose of the Study:

• To develop an accurate multistep prediction model for BTP in the sintering process.
• To address the limitations of traditional models in capturing complex spatial-temporal dynamics.
• To enhance the reliability and efficiency of sinter ore production.

Main Methods:

• Proposed a probabilistic spatial-temporal aware network (BTPNet) for BTP multistep prediction.
• Encoder network utilizes multichannel temporal convolutional network (MTCN) for temporal feature extraction.
• Introduced a variables interaction-aware module (VIAM) for spatial feature extraction.
• Decoder network incorporates probabilistic estimation (PE) to mitigate accumulated prediction errors.

Main Results:

• BTPNet effectively extracts spatial-temporal features crucial for accurate BTP prediction.
• The proposed VIAM and PE methods enhance the model's predictive performance.
• Experimental results on a real sintering process show BTPNet outperforms existing state-of-the-art models.

Conclusions:

• BTPNet offers a robust solution for accurate BTP multistep prediction in industrial sintering.
• The model's ability to handle spatial-temporal dependencies improves process control and operational stability.
• This approach contributes to optimizing sinter ore quality and production efficiency.