Hello,
Check this thread, you'll probably find the solution.
Previously, I had similar problems that error rates didn't converge during DADA2-denosing and pair-ended reads didn't merge. After removing the primer sequence in the reads, the problems were fixed.