In this study, to classify Power Quality (PQ) disturbances, attributes are extracted by 2D Discrete Wavelet Transform (2D-DWT) method and Support Vector Machines, Artificial Neural Networks and Bagged Decision Trees (BDT) methods are used for classification stage. 2200 signals are synthetically produced for 11 different PQ disturbances, including noisy (40 dB, 30 dB and 20 dB) and noiseless states. Signals are transformed into 2D image matrices and 2D DWT is applied to each. Attributes are created by applying different level of decomposition and statistical properties. The most appropriate ones are selected with Sequential Forward Selection (SFS) and ReliefF methods. BDT method, which uses selected attributes with SFS, is the method that gives the best performance with a rate of 99.12 +/- 0.12%.