According to the problem that the particle swarm optimization(PSO) is easy to fall into local convergence, a new method named GPSO algorithm is proposed to improve the algorithm, which is based on the reaction factor and the law of gravity. The algorithm uses the gravity law to quickly determine the optimal direction of the particles. When the particles fall into local optimum, the particles are drapped out of local optimum. The simulation experiments show that the improved algorithm has achieved remarkable results in the convergence speed and the optimization ability. Finally, the model of ethylene cracking conversion was obtained by using the improved algorithm to optimize the parameters of BP neural network. The experimental results show that the neural network model based on the improved algorithm can better predict the conversion rate of ethylene.