Transmission Control Protocol Friendly Rate Control (TFRC), as a mechanism to implement User Datagram Protocol (UDP) congestion control, can share bandwidth resources friendly with Transmission Control Protocol (TCP) data streams. However, since it is still a connectionless transmission mode in nature, there are problems that congestion has occurred when the transmission rate is adjusted, and the curve changes too much when the channel is unstable. In order to improve these problems, this paper proposes a method of network change trend prediction based on Long Short-Term Memory (LSTM). This method collects the packet loss rate in a certain period of time for neural network training, predicts the packet loss rate at a later time according to the input packet loss rate data, and converts the predicted data into the packet loss rate change trend to join the algorithm. The simulation experiment results show that the output curve with prediction has smaller change interval and smoother change trend compared with the output curve of the original algorithm.