Particle Swarm Optimization with Subpopulations Shuffled Dynamic and Its Application in Soft-Sensor of Acrylonitrile Yield

This paper presents a variant of Particle Swarm Optimization, called Subpopulations Shuffled Dynamic of Particle Swarm Optimization (SSDPSO). In SSDPSO, particles are partitioned into different subpopulations by fitness to maintain diversity of population efficiency. The subpopulations will be shuffled together to be a new population after they evolved for certain iterations. Furthermore, the iterations which subpopulations evolve will be dynamic changing. Some of particles will be re-randomized when subpopulations stagnate for certain iterations. A portion of shuffled population with poor position will be substituted by other one with better position. The performance of SSDPSO is investigated by some benchmark functions and compared with other version PSO. The results show that SSDPSO can achieve better solutions and get faster convergence. SSDPSO is then applied to train artificial neural networks to construct a soft-sensor of acrylonitrile yield. The results show that the soft-sensing model constructed by SSDPSONN is feasible and effective.