I have seen a similar effect myself, and the exact cause is still a bit of a mystery.
But what I did was connect a diode in series with each series string of panels before the outputs are combined, and that seemed to fix it.
I suspect it might have something to do with partial shading of one or more panels. The voltage in one string falls, and that loads down the other string.
The mppt software will still heroically try to find the power peak, but it pulls the voltage right down in the process.
With diodes fitted everything then worked as expected. The power peak coincided pretty well with the voltage on the rating plate.
There may be some other mysterious mechanism at work, but diodes certainly solved the problem here.
Anyhow, its worth a try.