fMRIPrep will, by default, use an algorithm to detect dummy scans. If you know exactly how many there are, you can override this. This avoids situations where some runs are estimated to have more dummy scans than others, or simply allows you to generate regressors that fit your analytical needs, when you would drop more volumes than the algorithm would detect.
The main places this has an impact are slice-timing correction and CompCor. Slice-timing correction will leave dummy scans untouched, to avoid polluting the first non-dummy volume with non-steady-state effects. CompCor will ignore the dummy volumes to avoid giving you regressors that capture non-steady-state effects that will be accounted for by the non_steady_state_XX regressors.