The present fault early-warning models for hydropower equipments are mostly small and scenario-customized models with low reusability, which places a significant burden on model development, operations and maintenance. A hydropower equipment fault early warning method based on time series foundation models was proposed, aiming at a general early warning goal of one model for multiple scenarios. To address the dilemma that time series foundation models struggle to effectively distinguish between normal and abnormal data due to their strong cross-distribution generalization ability, a similarity-constrained framework based on latent similarity-based modeling was novelly developed for time series foundation models. The framework constructs similarity-based constraints between the representations of test data and prototypes, thus limiting the model's performance for abnormal data. The theoretical analysis results and several real-world equipment fault case analysis results from a large hydropower station in the Yangtze River basin indicate that, the framework is capable of enlarging the prediction error gap between normal and abnormal data, improving the identifiability of abnormal patterns, and improving accuracy and earliness of early warning. The framework provides a practical approach for application of time series foundation models for general hydropower equipment early warning.
YANG Jing-yu, KONG Li-jun, TIAN Yuan, LIU Dong, ZHANG Zhi, TANG Zheng-yang.
General Early Warning Method for Hydropower Equipments Based on Similarity-constrained Time Series Foundation Models[J].
Science Technology and Engineering, 2026, 26(31): 13536-13545 DOI:10.12404/j.issn.1671-1815.2507999
作为时序分析领域的人工智能技术前沿,时序基础模型(time series foundation models, TSFM)凭借强大的表征泛化能力展现出跨场景、跨分布应用的巨大潜力[13-15]。TSFM通过在海量、多源时序数据上进行自监督预训练来学习通用时序特征表示,训练得到数亿乃至数十亿参数规模的模型,能够捕捉不同场景的共性动态模式,实现跨场景、跨分布的泛化推理。因此,TSFM有望大幅降低水电设备预警模型的任务定制成本,支持实现“一模型多场景”的通用预警目标。
依据机组的有功功率对数据进行稳态过滤,仅使用稳态数据进行训练与测试。对数据进行归一化处理,将输入步长和预测步长分别设定为512和64。采用Time-MoE的开源版本作为模型基座,其总参数规模为4.53亿,隐特征维度为768。原型特征的数量设定为384,组合权重稀疏分位数设定为95%。如图5所示为SC-TSFM在三个案例中的测试数据平均预测误差比值对于原型特征数量的敏感度分析结果,从中可以看出,SC-TSFM的平均预测误差比性能在一定范围内对于原型特征数量的设定不敏感,其主要原因在于组合权重稀疏化的使用。微调过程采用学习率为5×10-5的AdamW优化器,批尺寸和迭代回合数分别设定为256和10。采用阈值超越法(peak over threshold,POT)[20]确定阈值,该方法基于极值理论,根据异常评分的尾部分布自适应确定阈值,相比传统基于标准差的方法具备更高的统计合理性。采用累计越限报警策略,即仅当测试数据流的预测误差在一定时间内超过阈值的次数累计突破某数值时才会报警,从而有效避免偶发误差的负面影响。
WuQiuwei, ZhaoWenyan. Research overview of digital twin and generative artificial intelligence enabled new power system[J]. Electric Power Automation Equipment, 2025, 45(11): 59-71.
WuTing, WenShulin, YanZhaoli, et al. Unsupervised anomaly detection of hydro-turbine generator acoustics by integrating pre-trained audio large model and density estimation[J]. Journal of Electronics & Information Technology, 2026, 48(2): 772-783.
GongYouhua, XinWei, YuQiang, et al. Wind turbine gearbox fault diagnosis based on multimodal feature fusion of acoustic patterns[J]. Science Technology and Engineering, 2026, 26(5): 1961-1971.
GaoXuke, QinChao, GaoFeng, et al. Dual parameter early warning model for winding faults in distribution dry-type transformers based on LSTM-GMM[J]. Science Technology and Engineering, 2024, 24(31): 13407-13414.
PiYouchun, TanYun, GuoYujing, et al. Vibration fault warning of hydroelectric generating sets based on improved VMD and GRU[J]. China Rural Water and Hydropower, 2024(3): 244-249.
LuoYun, LiZhanguo, FuLongxia. Research on fault warning of coal mill based on feature selection and multivariate regression prediction with BiLSTM[J]. Journal of Chinese Society of Power Engineering, 2025, 45(5): 724-732.
WangYongfei, LiXiaofei, SunYuxin, et al. A monitoring and warning method for hydroelectric units based on KPCA and LSTM[J]. Journal of Vibration and Shock, 2024, 43(24): 287-294.
[15]
LiuZ G, ZhangH, ZhouY, et al. Machine learning-based early warning system for abnormal oil levels in hydroelectric power station pressure oil systems[C]// Ninth International Conference on Energy System, Electricity, and Power. Bellingham: SPIE, 2025: 1172-1178.
LiTeng, LiaoJun, FanPeipei, et al. Forecasting method for ultra-high voltage converter transformer top-oil temperature based on self-attention mechanism and improved recurrent neural network hybrid model[J]. Modern Electric Power, 2024, 41(6): 1167-1175.
[18]
胡航宁. 基于深度学习的水电设备异常检测研究[D]. 哈尔滨: 哈尔滨工业大学, 2023.
[19]
HuHangning. Research on anomaly detection of hydropower plant equipment based on deep learning[D]. Harbin: Harbin Institute of Technology, 2023.
XuLiyan, XuKang, HuangXingting, et al. Transformer-based method of anomaly detection for time series data[J]. Computer Technology and Development, 2023, 33(3): 152-160.
LiChen, LinWei, XuLi. TCN-GAN-based temporal traffic anomaly detection[J]. Journal of Information Security Research, 2025, 11(10): 907-916.
[24]
AnsariA F, StellaL, TurkmenC, et al. Chronos: learning the language of time series[J]. Transactions on Machine Learning Research. arXiv Preprint, arXiv: 2403.07815, 2024.
[25]
GoswamiM, SzaferK, ChoudhryA, et al. MOMENT: a family of open time-series foundation models[C]//International Conference on Machine Learning. Vienna: PMLR, 2024: 16115-16152.
[26]
ShiX M, WangS Y, NieY Q, et al. Time-MoE: billion-scale time series foundation models with mixture of experts[C]//International Conference on Learning Representations. Singapore: ICLR, 2025: 1-33.
[27]
GuoD Y, YangD J, ZhangH W, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning[J]. Nature, 2025, 645(8081): 633-638.
[28]
KirillovA, MintunE, RaviN, et al. Segment anything[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. New York: IEEE, 2023: 4015-4026.
[29]
YaoQ, YangC H H, JiangR, et al. Towards neural scaling laws for time series foundation models[C]//International Conference on Learning Representations. Virtual Event: ICLR, 2025: 1-28.
[30]
AlmeidaN M, CamargosM O, MarianoD G B, et al. An evolving approach to the similarity-based modeling for online clustering in non-stationary environments[J]. Evolving Systems, 2025, 16(1): 16.
[31]
SifferA, FouqueP A, TermierA, et al. Anomaly detection in streams with extreme value theory[C]//Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2017: 1067-1075.