An example might help you to understand the concept ...
In the time domain, if you know that your signal is well approximated locally by a polynomial of degree D, then you would fit that polynomial to the measured waveform, over a finite analysis window. The signal power would be the power of the fitted polynomial (or the sum of squared polynomial coefficients), whereas the noise power would be the sum-of-squared residuals.
Alternatively, if your signal is better modelled as a sinusoid, then use a set of complex sinusoids that are orthonormal over the analysis interval instead of the polynomial. Of course, this is done most efficiently in the frequency domain, using the FFT.