I'm not able to inspect your model at the moment but there's a few things I notice about your problem. From the graphs I'm assuming you're creating a device level detailed switching model of the inverter, with coupled thermal effects?
There's a ton of lookup table work going on here, and if it's trying to solve the switching transitions, then you're going to have a lot of very small but important dynamics relative to your timescale. This creates what's known as a stiff problem. Try using a different solver like ODE15s rather than 23t. Also consider a fixed step with a very small time step as it might capture the dynamics sufficiently with fewer minor step.
Modeling detailed device performance can be resource intensive, but even so there aren't that many devices in this problem.
Do you need the detailed thermal coupling for the transient start-up? If not it might be easier to consider the losses only for a steady state model of the switching.
The errant results you're seeing for 800 volts looks like possibly something in a lookup table isn't extrapolating properly.
