A novel SVD-UKFNN algorithm for predicting current efficiency of aluminum electrolysis.

Area of Science:

• Chemical Engineering
• Computational Science

Background:

• Aluminum electrolysis production process (AEPP) efficiency is vital for industrial output.
• Dynamic nonlinearity and complexity in AEPP challenge accurate prediction models.

Purpose of the Study:

• To develop an advanced model for predicting AEPP current efficiency.
• To enhance prediction accuracy and robustness in complex industrial systems.

Main Methods:

• Constructed a dynamic prediction model using an artificial neural network (ANN) within an unscented Kalman filtering neural network (UKFNN).
• Integrated singular value decomposition (SVD) with UKFNN for improved numerical stability.
• Redefined prediction variance as a cost function optimized via gradient descent for reduced error accumulation.

Main Results:

• The proposed NSVD-UKFNN significantly improved prediction accuracy.
• Achieved a 2.08-fold reduction in mean absolute error (MAE).
• Reduced sum of squared errors (SSE) by approximately 22.23 times compared to baseline models.

Conclusions:

• The NSVD-UKFNN offers a robust and accurate solution for predicting current efficiency in AEPP.
• This method addresses the challenges posed by AEPP's inherent complexity.
• The findings support enhanced industrial production efficiency and quality in aluminum electrolysis.