in principle, with a standard supervised feedforward ann, you do not _need_ to normalize the inputs : the training process will adapt the input weights to the scale of your data ; it will however take time and you can hope to speed-up the training by normalizing your inputs in a reasonable manner (that is a manner which takes the distribution into account : for instance, blind min-max normalization may hurt your training badly if you have a very large outlier)
the only thing you _must_ to normalize is your target data so that it fits in the range of your output layer function !
You don't have a problem doing that. The first 20 variables are using a common measure in the range [0,1]. Put the last variable in the same range by using any membership function of your choice, even an inverse function can do that.
Any book on fuzzy logic will do. Normalization is the basis for fuzzy computations and allows you to combine data in the same scale. Everything depends on your knowledge engineering skill to ferret out information from experts on the way they solve their problem. You work with experts in the area of your application to determine the threshold of your data.