Various types of distributed equipment and intelligent equipment are connected to the power system. It makes the power system more and more sensitive to power fluctuations, which has led to the identification and processing of Power Quality Disturbances (PQD) become increasingly important. By combining the Segmented Modified S-Transform (SMST) and the Random Forest (RF) algorithm, a new method for PQD identification under complex noise conditions is proposed. Firstly, different frequency bands of SMST are tuned based on various detection errors and kurtosis, and 75 time-frequency features are extracted from the signal using SMST to form the original feature set. Then, the node splitting process of Classification Regression Tree (CART) is improved. The discrete value processing strategy is added and the drop of Gini index is used as the new node splitting rule. Moreover, before the next node splitting, the feature whose Gini index drops to zero is removed. Finally, RF classifier is constructed with modified CART algorithm and used to classify the complex PQD signals. Experiments show that under the condition of different SNR, the new method can effectively identify most single PQD signals and common dual-compounded PQD signals. Although the new method still has some room for improvement in terms of efficiency, its improvement at different aspects can effectively benefit the accuracy of PQD recognition, and its average classification accuracy is significantly higher than traditional PQD recognition methods based on S-transform. This work is supported by National Natural Science Foundation of China (No. 61603212 and No. 51407104).