Predicting Earth Rotation Parameters (ERP) is crucial for the precise positioning and navigation both on Earth’s surface and space. As complex variations of the Earth's rotation, ranging from high-frequency trembles, inter-annual to -decals oscillations, however, high-precision ERP predictions are rather challenging. For the accurate predictions of these stable signals, we develop an simple, adaptive yet high-precision HSA + AR model improved in muti-scales frequencies: (1) Hankel-z quarter-parameters {A, α, f, θ} (amplitude, damping, frequency, phase) harmonic fit of Chandler, Annual/Semi-annual wobbles over the past 10-years; (2) Multiple-peaks low-frequency fit over the 1962-now; both of step (1)-(2) termed as HSA method and (3) AR modelling of the irregular residual variations. In the secular stability test, our results derived from final IERS C04 ERP solutions can largely reduce the forecast errors beyond 60% in each ERP components (within 1–90 days), compared the LS + AR methods. Considering the superior short-term calibration by ERP high-frequency terms and surface fluid excitation, we determine the rapid GNSS ERP (HSA (rapid) + AR, low latency) and GFZ EAM solutions (HSA (rapid) + AR (EAM)) as predictive ERP basis. In the real experiments, HSA (rapid) + AR (EAM) in short-term (1–20 days) and HSA (rapid) + AR in middle-term (20–90 days) predictions significantly surpass the benchmark from the official IERS Bulletin A predictions and mostly studies in the same predictive epochs. Both rapid ERP and EAM solutions are featured with a shorter latency of 1–2 days. Therefore, our strategy further advanced the real-time ERP predictions greatly compared to the existing predictive solutions, which support various scientific and operational applications in future.