Accurate detection of abnormal transient feedwater flow rates is critical for ensuring the safe operation of multi-module high-temperature gas-cooled reactors (HTGRs). While traditional deep learning models are widely used for fault diagnosis in nuclear power plants (NPPs), these models often lack interpretability and fail to pinpoint the exact locations of faults within complex systems. Building on our previous research, this paper proposes a novel deep learning architecture that integrates Shapelet with Graph Convolutional Networks (GCNs) to address these limitations. By replacing the Long Short-Term Memory (LSTM) network with Shapelet for time-series feature extraction, our model not only achieves a 100% detection rate for abnormal feedwater flow on the original dataset, but also extracts interpretable features that reveal statistically significant patterns, such as rising or falling trends. Additionally, the node importance derived from the GCN enables precise localization of the specific nodes responsible for abnormal feedwater flow, enhancing both diagnostic accuracy and interpretability. Experimental results demonstrate the effectiveness of the proposed approach in detecting and localizing transient feedwater flow anomalies. Compared to our previous LSTM-GCN model, the Shapelet-GCN model offers a more robust and interpretable solution for fault diagnosis in HTGRs, contributing to improved safety and operational efficiency in NPPs.