Abstract: In order to solve the problem of inaccurate measurement caused by excessive stress response movement during weighing, a flock of sheep dynamic weighing system was constucted and a dynamic weighing method of sheep with improved GA-PSO-BP neural network was designed, that is, the moving average filtering algorithm was used to process the interference data smoothly, and the improved GA-PSO hybrid algorithm was used to optimize the weights and thresholds of the BP neural network. The three models, namely BP neural network, PSO-BP neural network and improved GA-PSO-BP neural network, were trained and compared in performance. The results showed that the improved GA-PSO-BP neural network had the highest prediction accuracy, and the weight value with the prediction error within 1.00 kg accounted for 85%. The average absolute percentage error of the improved GA-PSO-BP neural network was 0.679 6%, which was higher than that of BP neural network（1.007 4%） and PSO-BP neural network（0.975 2%）, but the time required to measure a single sheep was 11.2 s, which was slightly lower than that of PSO-BP neural network（10.5 s）. The results indicated that the dynamic weighing method of sheep based on improved GA-PSO-BP neural network could meet the requirements of intelligent feeding in pastures.