The length of day (LOD), a crucial component of Earth orientation parameters (EOP), arises from fluctuations in Earth's rotation rate due to internal and external forces. These variations manifest as increases or decreases in LOD, directly influencing the timescale of the diurnal cycle. This study employs five distinct methods—least squares auto regressive (LSAR), weighted least squares auto-regressive (WLSAR), long short-term memory (LSTM) combined with polynomial curve fitting (PCF) extrapolation and least squares (LS) extrapolation, a hybrid LSTM and LS model (LSTM+LS), and a hybrid LSTM and weighted least squares model (LSTM+WLS), corresponding to schemes 1 to 5 in this study—to predict the LOD time series from January 1, 2016, to December 31, 2020, based on the EOP 20 C04 dataset released by the International Earth Rotation Service (IERS). The proposed scheme 5 (LSTM+WLS) in this study involves applying WLS method to the LOD data corrected for solid Earth zonal tidal effects to derive extrapolated, fitted, and residual terms. The residual term is then predicted using an LSTM model incorporating effective angular momentum (EAM) data. Finally, the LOD predictions are obtained by combining the predicted residuals, extrapolated terms, and solid Earth zonal tidal corrections. Compared to the other four schemes, scheme 5 demonstrates superior performance in 10-day predictions, achieving a mean absolute error (MAE) of 0.127 3 ms, representing improvements of 5.7%, 5.0%, 2.6%, and 4.6%, respectively. For 30-day predictions, it slightly outperforms schemes 1 and 2 while performing comparably to Schemes 3 and 4. In 90-day predictions, the MAE reaches 0.167 0 ms, with improvements of 8.0%, 8.8%, 15.3%, and 13.3% over the other schemes. Overall, the proposed LSTM+WLS model exhibits excellent performance in short-term LOD forecasting.
