choose attribute help

[weka77]

I use MultilayerPerceptron weka's classifier to do prediction. I'm using cross-validation. It gives me good results (98%) using train data. The problem is when I make prediction with the model obteined: the dataset that I need to Predict has a lot of attribute always null (while during the training these are not null).
Could It be a reason to obtain wrong percentage?
If I haven't values for a lot of attributes in dataset to predict I have to put these attribute off training set too?
how can I choose the right attribute?
thanks