In this paper, we propose a Packet Cache-Forward(PCF) method based on improved Bayesian outlier detection to eliminate out-of-order packets caused by transmission path drastically degradation during handover events in the moving satellite networks, for improving the performance of TCP. The proposed method uses an access node satellite to cache all received packets in a short time when handover occurs and forward them out in order. To calculate the cache time accurately, this paper establishes the Bayesian based mixture model for detecting delay outliers of the entire handover scheme. In view of the outliers’ misjudgment, an updated classification threshold and the sliding window has been suggested to correct category collections and model parameters for the purpose of quickly identifying exact compensation delay in the varied network load statuses. Simulation shows that, comparing to average processing delay detection method, the average accuracy rate was scaled up by about 4.0%, and there is about 5.5% cut in error rate in the meantime. It also behaves well even though testing with big dataset. Benefiting from the advantage of the proposed scheme in terms of performance, comparing to conventional independent handover and network controlled synchronizedhandover in simulated LEO satellite networks, the proposed independent handover with PCF eliminates packet out-of-order issue to get better improvement on congestion window. Eventually the average delay decreases more than 70% and TCP performance has improved more than 300%. In this paper, we propose a Packet Cache-Forward (PCF) method based on improved Bayesian outlier detection to eliminate out-of-order packets caused by transmission path drastically degradation during handover events in the moving satellite networks, for improving the performance of TCP The proposed method uses an access node satellite to cache all received packets in a short time when switching occurs and forwards them out in order. In view of the outliers’ misjudgment, an updated classification threshold and the sliding window has been suggested to correct category collections and model parameters for the purpose of quickly identifying exactly compensation delay in the varied network load statuses. to average processing delay detection method, the average accuracy rate was scaled up by about 4.0%, and t here is about 5.5% cut in error rate in the meantime. It also behaves well even though testing with big dataset. Benefiting from the advantage of the proposed scheme in terms of performance, comparing to conventional independent handover and network controlled synchronizedhandover in networks, the proposed independent handover with PCF eliminates packet out-of-order issue to get better improvement on congestion window. Eventually the average delay decreased more than 70% and TCP performance has improved more than 300%.