A Parallel Encoder Transformer with Multiscale Convolution for Industrial Temperature Prediction in Rotary Kiln

Predicting the temperature in rotary kilns is critical for ensuring product quality. However, traditional prediction methods often struggle with capturing the complex, multiscale temporal dependencies. To address this limitation, we propose a Parallel Transformer model specifically for rotary kiln temperature prediction. First, the proposed model proposes a Separable Multiscale Convolution (SMC) block that efficiently extracts multiscale features with significantly reduced computational complexity. Then, a dual-channel encoder structure is designed, which independently processes temporal and variable dependencies through parallel self-attention mechanisms. Our experimental results demonstrate a significant improvement in temperature prediction compared to existing state-of-the-art methods.