Aiming to address the challenge of low model accuracy in traditional benzene-toluene-xylene (BTX) yield prediction, this work proposes a novel prediction approach based on a pyramid attention mechanism (PAM) combined with a gated recurrent unit (GRU), referred to as the PAM-GRU model. The PAM of the proposed method can enable the construction of a hierarchical attention structure for multi-feature sequential data, allowing for the extraction of spatial features. In parallel, the GRU is utilized to capture dynamic temporal information within the time-series data through its recurrent structure, enabling the proposed prediction method to uncover underlying temporal variation patterns. By seamlessly integrating both spatial and temporal features, the PAM-GRU method achieves a more accurate and reliable prediction. Finally, the proposed method is applied to a real-world continuous reforming chemical production process, and the proposed model performance is evaluated using various metrics. The proposed PAM-GRU model is compared with models such as the recurrent neural networks (RNN), the long short-term memory networks (LSTM), the gated recurrent units (GRU), the LSTM based on attention mechanisms (Attention-LSTM), and the GRU based on attention mechanisms (Attention-GRU) in terms of four indicators: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The results show that the proposed PAM-GRU model can effectively integrate the spatial and temporal features in the continuous reforming production process, achieving efficient and accurate prediction of the BTX yield. In addition, to cope with the complex production environment, a robustness test is added in the proposed model. The results show that the proposed model has strong robustness and can effectively suppress the interference of sudden noise.