At present, most of the dynamic prediction methods of ice storage air conditioning cooling load showed poor precision and slow convergence speed, due to bad correlation between model input variables and output results, high information redundancy. This paper proposes an improved PSO-BP neural network optimization algorithm to predict the cooling load of large public buildings. For the input variables and the output results, the degree of grey correlation analysis is used to eliminate the chance of the logarithm of the samples input variables, and the key factors affecting the cold load of the ice storage air conditioning system are determined as input variables to predict the dynamic cooling load of the ice storage air conditioning system. The results show that the key factors of the cooling load of ice storage air conditioning system include the outdoor air temperature at T h, the outdoor air temperature at T-1 h, the outdoor air humidity at T h, the solar radiation intensity at T h, the solar radiation intensity at T-1 h, and the air-conditioning cooling load at T-1 h. Therefore, these key factors act as the input variables of the prediction model which show higher accuracy and faster convergence speed than the traditional PSO-BP network model of full-variable prediction method.