A self-supervised LSTM network for cell temperature prediction in aluminium electrolysis reduction

These paper presents a novel LSTM method for cell temperature prediction in aluminum electrolysis industry. In the practical electrolysis production, it always suffers the limitation of the sample scarcity, which greatly decrease the performance of the supervised-based prediction models. This paper proposes an enhanced LSTM unit which fully utilizes the self-supervised loss as the training process. The Kullback-Leibler (KL) divergence is proposed for learning the similarity of the different unlabeled data. The proposed SSLSTM is further applied to the industrial aluminum electrolysis temperature process. The experimental results demonstrate that the proposed method has the state-of-the-art accuracy and robustness.