Quantification of abnormal characteristics and flow-patterns identification in pumped storage system
Fei Chen,Zhigao Zhao,Xianghui He,Xiaoxi Hu,Jianlin Chen,Pan Liu,Xiuxing Yin,Jiandong Yang
DOI: https://doi.org/10.1007/s11071-024-10131-x
IF: 5.741
2024-10-12
Nonlinear Dynamics
Abstract:The pumped storage units (PSUs) deviate from the optimal operating condition, and the abnormal flow pattern generated in the draft tube seriously affects the safe and stable operation of the power station. Precise identification of abnormal flow patterns in the draft tube is an effective measure to improve the energy conversion efficiency of PSUs. Existing data-driven methods for identifying abnormal flow patterns in hydraulic turbines primarily focus on classification accuracy, disregarding the fact that abnormal flow regimes are rare events. How to detect and identify the abnormal flow pattern of PSUs from vast operational data is a key challenge to be addressed in the real power plant. Therefore, this paper builds a four-stages monitoring framework that integrates signal acquisition, feature extraction, anomaly detection, and anomaly identification for operational flow patterns in PSUs. In Stage 1, a dataset of different operational flow patterns of the draft tube is formed based on the PSU and the data monitoring system. In Stage 2, a new nonlinear dynamics metric named modified multiscale time and frequency indicator (MMTFI) is proposed to quantify the operational characteristics of different abnormal states of the draft tube from both the time and frequency domains. In Stage 3, the anomaly detection module for abnormal flow patterns within the draft tube is constructed based on normal operation data, based on MMTFI and support vector data description (SVDD), to achieve accurate capture of abnormal flow patterns from massive data. In the stage 4, an anomaly identification module for abnormal flow patterns within the draft tube is constructed based on MMTFI and adaptive kernel extreme learning machine (AKELM) to achieve accurate identification of anomalous flow patterns. Its core lies in the integration of two artificial intelligence paradigms, unsupervised learning and supervised learning, to realize the detection and identification of limited abnormal flow pattern signals from massive data. The framework is applied to a micro PSU, and the experimental results demonstrate that the proposed framework accurately detects anomalous flow patterns, achieving an identification accuracy of 95.10%, which demonstrates a more outstanding performance in comparison with other models. It will provide a powerful tool for reducing the probability of fault incidents and improving energy conversion efficiency in PSUs.
engineering, mechanical,mechanics