Major power outages both domestically and internationally highlight that the key transmission sections (KTS) represents a vulnerable link within the interconnected power grid. Hence, Rapid and accurate diagnosis of weaknesses and faults in the power grid is beneficial for the dispatch department to make timely adjustments to the operational modes. However, existing fault detect algorithms are seldom tailored to KTS. Therefore, this paper presents a new method for the classification and detection of power grid KTS faults based on deep learning, using a gated recurrent unit (GRU) network model with Nadam optimizer. First, fault sample data can be acquired through batch simulation and then converted into grayscale images. Next, a GRU-based fault recognition classifier is trained utilizing Nadam optimization. In addition, dropout techniques and self-assessment mechanisms have been incorporated into the framework. The findings indicate that the GRU network optimized with Nadam achieves the highest accuracy rates of 99.90, 99.87, and 99.63%. Furthermore, its performance remains unaffected by variations in fault location and onset time.