【摘要】 在铝电解生产过程中，为保持电解槽的能量平衡与物料平衡，需依据专家经验向生产系统提供控制参数确保槽况稳定。但铝电解槽历史生产数据有着积累过长、参数相关性复杂等问题，仅凭专家经验无法准确判断需进行参数调节的铝电解槽。针对上述问题，本文设计一种基于注意力机制的铝电解槽多元时序辅助决策方法：利用空间编码器代替卷积网络提取局部特征，建立自注意力编码器神经网络多元时序分类模型，同时融合部分专家的历史经验，提高模型对待调节铝电解槽判别的准确率。消融实验的结果表明空间编码器模块优于CNN模块；在某铝厂的真实生产数据上的实验结果表明，该方法优于专家经验、DTW等传统方法，在铝电解实际生产决策中具有使用价值。更多还原

【Abstract】 In the process of aluminum electrolysis production, in order to maintain energy balance and material balance of the electrolytic cell, it is necessary to provide control parameters to the production system based on expert experience to ensure stable cell conditions. However, the historical production data of aluminum electrolysis cells have problems such as excessive accumulation and complex parameter correlation, and relying solely on expert experience cannot accurately determine the aluminum electrolysis cells that require parameter adjustment. In response to the above issues, this paper designs an attention mechanism-based multi temporal auxiliary decision-making method for aluminum electrolysis cells: using spatial encoders instead of convolutional networks to extract local features, establishing a self attention encoder neural network multi temporal classification model, and integrating the historical experience of some experts to improve the accuracy of the model in discriminating aluminum electrolysis cells. The results of the ablation experiment indicate that the spatial encoder module is superior to the CNN module; The experimental results on real production data of an aluminum smelter show that this method is superior to traditional methods such as expert experience and DTW, with practical value provided in decision-making of aluminum electrolysis production.更多还原