Improved Particle Swarm Optimization-GPR for Alumina Production Index Prediction Model

This paper addresses the issue of excessive reliance on initial values and susceptibility to local optima when using Gaussian process regression (GPR) to predict comprehensive production indices in the alumina production process. Fur-thermore, the authors proposed an improved particle swarm optimization Gaussian process regression (IPSO-GPR) algorithm. Firstly, a population initialization strategy was introduced by integrating logistic chaos mapping with an elite dynamic reverse learning mechanism to improve the quality of initial populations while accelerating convergence rates. Secondly, we proposed an adaptive velocity update method with dynamic inertia weight coefficients and learning factors to balance the global population and local search capabilities of the algorithm. This strategy enabled particles to dynamically adjust their search direction and speed during the optimization process, thereby preventing the population from falling into local optima. Lastly, a mutation strategy incorporating across generations horizontal and vertical crossover was proposed to reduce search blind spots, enhancing population diversity and solution accuracy. Comparative experiments with various mainstream hyperparameter optimization al-gorithms on alumina comprehensive production index prediction tasks demonstrate that GPR optimized with IPSO-GPR exhibits superior fitting accuracy and generalization capability.