With gigahertz-level bandwidth, terahertz (THz) holds promise for achieving exceptionally high transmission rates in prospective sixth-generation (6G) communications. However, considerable loss poses an obstacle to THz communication. To compensate for this, massive multiple-input-multiple-output (MIMO) based beamforming is utilized to promote directional power with narrow beams in communication. In dynamic environments, the frequent adjustment of narrow beams results in fast time-varying channel state information (CSI), which constrains the application of the THz communication systems. While traditional deterministic-based and statistical-based channel tracking methods address different aspects of this issue, they suffer from balancing accuracy and complexity in the THz dynamic environments. To solve this problem, based on the cluster distribution of THz time-varying channel, we propose a novel hybrid channel tracking method that uses deterministic physical motion variation law to extract the cluster subspace, and then statistical Markov evolution models are applied within it. To achieve this, an integrated clustering and estimation method, clustering subspace matching pursuit (CSMP) is proposed for obtaining the channel clusters prior. Then based on above hybrid tracking method design, we propose a virtual cluster subspace turbo-approximate message passing (VCS-TAMP). Finally, several simulation results validate that our proposal achieves great improvement in both accuracy and computational time performance.