A multi-UAV system relies on network connectivity to carry out diverse missions effectively. However, due to their inherent mobility, multi-UAV networks often encounter highly dynamic network conditions, which increase packet loss and trigger frequent retransmissions. Without considering mobility, selecting an inappropriate packet length causes overhead due to headers and packet retransmission, ultimately impeding the efficiency of the mission. This paper proposes an LSTM-based Retransmission Prediction and Packet Length Optimization (RPPLO). This scheme predicts the length of packets likely to be retransmitted and then dynamically adjusts packet size to enhance overall network efficiency. During the mission, each UAV collects mobility features such as its own and neighbor velocity, positional features such as distances between UAVs, network metrics, including current end-to-end delay, and the length of recently transmitted packets. The LSTM model learns from features related to mobility and predicts the length of packets that are likely to be retransmitted under the current environment. If the length of retransmitted packets increases, RP-PLO proactively shortens the packet length; when the predictions indicate a decrease in the length of retransmitted packets in a stable network environment, the packet length is increased to improve transmission efficiency. By adopting the packet length to a dynamic network conditions, the proposed method achieves a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.54 \times$</tex> reduction in retransmission length compared to a fixed-length scheme with a similar transmission length, a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.15 \times$</tex> increase in transmission length compared to a scheme with a similar retransmission length with higher packet transmission efficiency.