Accurate power quality disturbance (PQD) classification is significantly important for power grid pollution control. However, the use of nonlinear loads makes power system signals complex and distorted and, thus, increases the difficulty of detecting and classifying PQD signals. To address this issue, this article first proposes an optimized S-transform (OST). It optimizes different window parameters to improve time-frequency resolution using maximum energy concentration. A kernel SVM (KSVM) classifier is proposed to classify multiple features using a combination of kernels. Integrating OST and KSVM, a classification framework is further proposed to detect and classify various PQD signals. Extensive experiments on computer simulations and experimental signals demonstrate that the proposed classification framework shows better performance than several state-of-art methods in classifying not only single and multiple PQD signals but also PQD signals with different noise levels. More importantly, our framework has superior performance in detecting nonlinearly mixed PQD signals.