COMPARATIVE ANALYSIS OF MACHINE LEARNING MODELS FOR ESTIMATING STEAM TURBINE ROTOR VIBRATION USING OPERATIONAL DATA FROM A NIGERIAN THERMAL POWER PLANT

Abstract

Steam turbine rotor vibration is an important condition indicator for rotating machinery in thermal power plants. This study compares Artificial Neural Network (ANN), k-Nearest Neighbours (KNN), Support Vector Machine (SVM), and Decision Tree (DT) regression models for contemporaneous estimation of a composite rotor vibration index from supervisory control and data acquisition records obtained from one steam turbine unit at Egbin Thermal Power Plant, Nigeria. The raw export contained 988 records collected from 1 January to 24 March 2021 at a nominal two-hour interval. After explicit data-quality checks and removal of low-load records, 929 observations were retained. The target was the root mean square of six vibration channels, while active power, reactive power, rotor temperature, and stator core temperature were used as predictors. Data were partitioned into training, validation, and held-out test subsets before z-score scaling, and all scaling parameters and the elevated-vibration screening threshold were derived without using the test set. On the held-out test set, KNN obtained the highest regression coefficient of determination (R² = 0.720) and the lowest RMSE (0.00365), followed by DT (R² = 0.673), ANN (R² = 0.669), and SVM (R² = 0.647). For the secondary elevated-vibration screening analysis, DT achieved the highest accuracy (0.950) and F1-score (0.829), while all four models produced AUC values between 0.926 and 0.939. These results support the use of data-driven models for rotor condition monitoring, while the single-unit dataset, zero-hour prediction horizon, and absence of plant sensor-location metadata limit claims about failure prevention, generalisability, and standards-based vibration severity.