I totally agree with the two contributors above .Just for the case that this should not resolve your issue one more proposal. You can weight your input data with the observed serial correlation pattern from your VECM yield autocorrelation problems and run the VECM with the weighted data again: of course you need to take the fact that you weighted the data into account when interpreting the results. This might even resolve the need to use a VECM so that you could eventually switch to a VAR. But that needs checking via cointegration tests.