A two-stage framework for adaptive boundary optimization in hydropower units using semi-supervised ensemble clustering and CNN-PatternNet risk assessment

The integration of variable renewable energy has intensified operational demands on large hydropower units for peak shaving and frequency regulation, forcing frequent operation in vibration-sensitive zones. This exacerbates mechanical wear and compromises system stability, highlighting the critical need for accurate, adaptive operational boundary delineation to enable safe scheduling. To overcome the limitations of existing static threshold-based methods, this study proposes a novel two-stage data-driven framework for adaptive boundary optimization. First, a semi-supervised optimized ensemble clustering model integrates three base clusterings through spatiotemporal consistency loss and particle swarm optimization, generating robust partitions of the operational space into prohibited, restricted, and stable regions. Second, a convolutional neural network-based pattern recognition model, trained on clustering-labeled data, quantifies operational risks and assesses high-risk duration and boundary fluctuations to provide quantitative support for boundary re-evaluation. Validated using real-world data from Ertan Hydropower Plant, the proposed clustering model delivers over 97% clustering accuracy and a silhouette coefficient of 0.92, identifying both documented and latent vibration zones while uncovering unit-specific boundary variations and relative threshold references. The risk recognition model achieves 99.8% classification accuracy on the historical labeled dataset, providing effective support for risk recognition and boundary consistency assessment. This framework offers an effective, adaptive solution to bolster hydropower units’ safety and reliability in renewable-integrated power systems, with important implications for intelligent dispatch and predictive maintenance.