Gas-insulated switchgear (GIS) exhibits various insulation defects during production and operation, and accurately identifying partial discharge signals caused by insulation defects is of significant importance for ensuring the safety of GIS equipment and power systems. By integrating the golden sine algorithm (golden-SA) to improve the subtraction-average-based optimizer (SABO), a fused golden sine-improved SABO optimization algorithm (GSABO) is obtained. This algorithm is applied to optimize parameters for the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and the kernel extreme learning machine (KELM) to achieve recognition of GIS partial discharge faults. Firstly, to address issues such as SABO possibly falling into local optima and insufficient convergence speed, chaotic mapping and the golden sine are introduced to improve it. Then, an experimental platform is set up to collect four types of typical partial discharge signals, which are decomposed using GSABO-ICEEMDAN, and effective modal components are screened using the correlation coefficient method. Finally, the sample entropy of the selected modal components is calculated to form a feature matrix, which is input into GSABO-KELM for fault classification and recognition. Experimental analysis shows that, compared to the unimproved SABO algorithm, GSABO demonstrates significant advantages in escaping local optima, convergence speed, and accuracy. Compared with other traditional algorithms, GSABO-ICEEMDAN-KELM achieves a recognition accuracy of 99.1667%, verifying the accuracy and superiority of this algorithm, which holds reference value for engineering applications in GIS partial discharge fault diagnosis.