Lid-driven cavity flow using two-relaxation-time LBM

Hi all,

Currently, I am trying to simulate a problem in which natural convection plays a rol due to a hot wall in a square cavity. Unfortunately, the model is unstable in most cases. In order to find the problem I simulated a lid-driven cavity, since it displays similar characteristics.
Now, when using SRT the flow is what one expects. However, after implementing TRT, the solution is very strange. The boundary nodes closest to the wall are given a velocity in the positive x-direction by bounce-back with a prescribed wall velocity. This is indeed seen in the solution. The row below the boundary nodes has a velocity with opposite sign. I ruled out an implementation mistake, as the TRT model with tau_s = tau_a (symmetric and anti-symmetric relaxation times) gives the SRT solution. The error is also happens for all relaxation times and gets worse when tau_s decreases (I used tau = 0.8 for the plots).
I starting to think it is an error inherent to TRT, but I am not able to find confirmation of such behaviour in the literature.

Does anyone have a solution or explanation for this behaviour? I will be very grateful.

Kind regards,