Fault Early Warning Method for Hydropower Units Based on an Improved Bees Algorithm with Two Parameters and Broad Learning System

To address pronounced fluctuations in monitored parameters and the difficulty of timely and accurate detection of incipient anomalies in hydropower units, this paper proposes a fault early warning method based on an improved Bees Algorithm with Two Parameters (BA2) and the broad learning system (BLS). First, while retaining the triangular distribution neighborhood search mechanism of BA2, opposition reflection search and a simulated annealing acceptance criterion are introduced to construct an improved BA2 (IBA2), and five-fold expanding window time series cross-validation is used as the fitness evaluation to automatically optimize key BLS hyperparameters. The optimized BLS is then applied to normal behavior modeling of thrust bearing temperature; operating variables under healthy conditions are used to predict thrust bearing temperature, thereby establishing a dynamic normal baseline adaptive to changes in operating conditions. On this basis, the model prediction residuals are processed using a 30-point moving average, and a fault early warning mechanism combining two-sided 3σ control limits with a persistent exceedance confirmation rule is developed to suppress false alarms caused by isolated noise and brief disturbances. Results from an actual unit show that the proposed method can identify the anomaly 381 min before the abnormality was recorded on site. Further comparisons with three prediction models show that IBA2-BLS achieves lower prediction errors and favorable stability across repeated runs. Benchmark function experiments indicate favorable optimization performance of IBA2, while ablation experiments support the effectiveness of the proposed optimization framework. These results demonstrate the feasibility of the proposed method for identifying incipient anomalies in the investigated hydropower unit and indicate its potential for condition monitoring and fault early warning applications.