中国有色冶金

In the aluminum electrolysis production process, the addition amount of aluminum fluoride plays a key role in maintaining the efficiency and stability of electrolysis. At present, it is mostly determined by experience. In fact, it has a complex nonlinear relationship with many factors such as alumina test data, aluminum level and electrolys temperature, showing dynamic changes, and it is difficult to make accurate decisions based on experience alone. This study addresses the challenges posed by the nonlinear, large time-delay, and strong coupling characteristics of the aluminum electrolysis process by integrating a soft attention mechanism into a BiLSTM network to construct a high-precision prediction model for aluminum fluoride dosage. Extensive data training, testing, and validation were conducted to ensure model reliability. The experimental results demonstrate that the proposed algorithm achieves exceptional prediction accuracy in estimating aluminum fluoride dosage. By leveraging the XGBoost algorithm to extract local features, the model enhances both prediction accuracy and operational efficiency. Furthermore, the integration of bidirectional LSTM enables the model to consider both forward and backward data dependencies, while the attention mechanism dynamically adjusts feature weights, further improving prediction performance. The XGBoost-BiLSTM-Attention model achieves an average error of 0.014, an average percentage error of 2.64%, and a linear fitting degree of 0.963, surpassing the overall performance of existing models. This prediction model provides significant decision-making support for precisely controlling aluminum fluoride dosage in aluminum electrolysis production, thereby enhancing production efficiency, reducing energy consumption, and achieving precise control.