Load-aware fault classification framework for distribution systems using k-means clustering and convolutional neural networks

• CNN framework for load-aware fault classification in distribution systems. • K-means clustering reduces model size and training complexity under varying loads. • EMD-BPF extracts stable features, generating time–frequency maps for CNN input. • Achieved 97.63% avg. accuracy, reducing training time from 30,678 s to 7,352 s. Accurate fault classification in distribution systems remains a challenge because load variability can impair the reliability and adaptability of conventional single-architecture models. In this study, a fault classification framework is proposed for distribution systems operating under various load levels using data generated through transient simulations of power systems. Fault scenarios were simulated, and pre-fault current data were clustered using k-means to categorise the load conditions, with each cluster assigned a dedicated convolutional neural network (CNN) sub-model. Input features were extracted via empirical mode decomposition (EMD), reconstructed using a band-pass filter (BPF), and converted into time–frequency energy maps. Without clustering, the proposed CNN-EMD-BPF framework achieved a high classification accuracy but required 30,678 s of training. With load-based clustering, the training time was reduced to 7,352 s while maintaining an accuracy of 97.63%. These results confirm that load-based clustering significantly reduces computational costs while preserving high accuracy, thereby enabling efficient and reliable fault classification in distribution systems.