An autonomous predictive regulator (Linear Quadratic Gaussian (LQG)) is being developed within the European H2020 ORP project, to be tested on the Gran Telescopio Canarias Adaptive Optics (GTCAO) system (about 400 actuators and 600 SH-WFS measurements). The system can run at 1kHz with 1.65ms loop delay in the predictive case, under DARC real-time computer. This LQG controller is based on a data-driven model identified using both machine learning and parametric identification techniques and automatically updated from telemetry data during operation. The complete procedure under OOMAO simulator is presented here (including calibration). The disturbance is generated thanks to StereoScidar measurements, with in addition turbulence bursts, windshake and vibrations. Specific mechanisms insure regulator stability. Performance in terms of Strehl ratio and stability is evaluated for both LQG and integrator regulators. These developments and results will serve the on-sky tests on GTC foreseen in 2024.