I want to identify fault and load patterns (signatures) from time series data of power measurements at distribution substations, principally via machine-learning and data mining and related approaches - i.e. SAX.
SAX may or may not be the best solution for this task. If you make an analogy with text, then SAX (esp SAX bitmaps) are good for long sentences, but for atomic words, you should use DTW or Euclidean distance.
DFT and DWT (and PAA, CHEB, PLA, APCA etc) are only to speed things up. Dont worry about them now.
Here is how you solve this problem.
Find ten examples of the same event in your space.
Plot them side by side.
Think about the invariances you would need to have an algorithm find them similar.
The invariances may include invariance to trend, offset, scale, uniform scaling, warping, phase etc (a good list is here http://www.cs.ucr.edu/~eamonn/Complexity-Invariant%20Distance%20Measure.pdf )
Once you understand the invariances you need, you can choose a normalization policy and a distance measure.
For example, maybe normalization policy = z-normalize and distance measure = euclidean distance