In the realm of data-driven power system risk assessment, the conventional use of convolutional neural networks is hampered by the issue of vanishing gradients, which in turn undermines the accuracy of predictions. This study proposes the integration of the transformer encoder neural network into the risk assessment framework for power systems, thereby enhancing the model’s ability to generalize. Furthermore, we introduce an innovative training data construction algorithm that leverages the cross-entropy technique. This algorithm is designed to augment the proportion of effective samples within the training dataset. Case study results demonstrate that our approach significantly boosts the predictive performance of the neural network model, as well as the precision of risk assessment in power systems.