Fang Xiaoyan, Fei Xihong, Wang Kang, Fang Tian, Chen Rui
School of Electronic and Electrical Engineering, Chongqing University of Science and Technology, Chongqing, 401331, China.
School of Electrical and Information Engineering, Anhui University of Technology, Ma'anshan, 243002, China.
Sci Rep. 2025 Mar 17;15(1):9173. doi: 10.1038/s41598-025-94210-y.
The prediction of current efficiency in the aluminum electrolysis production process (AEPP) is critical for improving industrial production efficiency and product quality. However, the inherent dynamic nonlinearity and multivariable complexity of AEPP hinder the development of accurate current efficiency prediction models. To address these challenges, a novel singular value decomposition unscented Kalman filtering neural network (NSVD-UKFNN) is proposed to improve the prediction accuracy of current efficiency in the AEPP. First, a dynamic prediction model is constructed within the framework of the unscented Kalman filtering neural network (UKFNN), employing artificial neural network (ANN) to capture the complex characteristics of the system. Second, singular value decomposition (SVD) is integrated with the UKFNN to compute the square root of the prior matrix, thereby improving the model's numerical stability. Finally, the prediction variance of state variables is redefined as a cost function and optimized using the gradient descent method to reduce error accumulation during the computation process, enhancing the prediction robustness of proposed method. The experimental results show that the proposed NSVD-UKFNN reduces the mean absolute error (MAE) by 2.08 times and the sum of squared errors (SSE) by approximately 22.23 times compared to the baseline model.
铝电解生产过程（AEPP）中电流效率的预测对于提高工业生产效率和产品质量至关重要。然而，AEPP固有的动态非线性和多变量复杂性阻碍了精确电流效率预测模型的发展。为应对这些挑战，提出了一种新颖的奇异值分解无迹卡尔曼滤波神经网络（NSVD-UKFNN），以提高AEPP中电流效率的预测精度。首先，在无迹卡尔曼滤波神经网络（UKFNN）框架内构建动态预测模型，利用人工神经网络（ANN）捕捉系统的复杂特性。其次，将奇异值分解（SVD）与UKFNN相结合，计算先验矩阵的平方根，从而提高模型的数值稳定性。最后，将状态变量的预测方差重新定义为代价函数，并使用梯度下降法进行优化，以减少计算过程中的误差积累，增强所提方法的预测鲁棒性。实验结果表明，与基线模型相比，所提NSVD-UKFNN将平均绝对误差（MAE）降低了2.08倍，将平方误差和（SSE）降低了约22.23倍。
