When can the standard error of the estimate be used to construct a prediction interval about a value y'?

You need the data to be homoscedastic, the errors to be independent. The independent variable(s) should lie within (or very close to) the range of observed values.