Abstract: To enhance the accuracy of TLE (Two-Line Element) in orbit prediction and address the issues of traditional polynomial fitting and physical modeling methods in dealing with the nonlinear evolution trend of orbits, a high-precision TLE orbit parameter fitting method based on Long Short-Term Memory (LSTM) networks is proposed. This method utilizes historical TLE data to conduct high-precision time series fitting and orbit synthesis of orbital elements. By applying LSTM neural networks to model the time series of TLE orbit parameters and combining it with polynomial fitting, a hybrid modeling strategy is formed. The experiment constructs a fitting model using the TLE data of 87 Iridium 33 debris and generates synthetic TLEs, using the SGP4 propagator to forward propagate the orbit position errors for three days. The experimental results show that the propagation error of the synthetic TLE is significantly reduced compared to the original TLE within three days, with an improvement rate of 97.70% on the third day; the propagation error of most targets is controlled within 2 km. The error of synthetic TLEs is concentrated and stable; the error frequency statistics show that the coverage ratio of the 0-2 km error range within three days exceeds 81%, which is significantly better than the original TLE. The TLE orbit synthesis method based on LSTM demonstrates superior performance in capturing the nonlinear evolution of orbit parameters. Combined with angle linearization and smoothing filtering strategies, it significantly improves the accuracy and stability of orbit parameter fitting. The research results have certain application value in enhancing space situational awareness, orbit collision warning, and mission scheduling.