Visualization of CNN Transient Voltage Classification Based on Feature Recognition and Enhancement

The deep learning algorithm represented by CNN has achieved remarkable results in complex feature extraction, classification and regression. Since the CNN adopts an end-to-end training process, the carrier of acquisition knowledge is a high-dimensional non-linear function, which is hard to understand by humans. Therefore, it is difficult to be widely used in online decision-making engineering scenarios such as the fast assessment of power system transient voltage stability. In this paper, the key input elements extracted by CNN are identified by Grad-CAM, and the key features captured by CNN in the assessment process are enhanced based on AM, so as to finally obtain the visual expression of key areas and variables of the original input data. Besides, a sample construction method based on single line diagram is proposed to improve the visualization effect of Grad-CAM and AM. Simulation results and analyses of IEEE 39bus test system demonstrates the effectiveness of the proposed method.