This paper investigates the start-up and shutdown phases of a five-bladed closed-impeller centrifugal pump through experimental analysis, capturing the temporal evolution of its hydraulic performances. The study also predicts the transient characteristics of the pump under non-rated operating conditions to assess the accuracy of various machine learning methods in forecasting its instantaneous performance. Results indicate that the pump’s transient behavior in power-frequency mode markedly differs from that in frequency-conversion mode. Specifically, the power-frequency mode achieves steady-state values faster and exhibits smaller fluctuations before stabilization compared to the other mode. During the start-up phase, as the steady-state flow rate increases, inlet and outlet pressures and head also rise, while torque and shaft power decrease, with rotational speed remaining largely unchanged. Conversely, during the shutdown phase, no significant changes were observed in torque, shaft power, or rotational speed. Six machine learning models, including Gaussian Process Regression (GPR), Decision Tree Regression (DTR), and Deep Learning Networks (DLN), demonstrated high accuracy in predicting the hydraulic performance of the centrifugal pump during the start-up and shutdown phases in both power-frequency and frequency-conversion conditions. The findings provide a theoretical foundation for improved prediction of pump hydraulic performance. For instance, when predicting head and flow rate during power-frequency start-up, GPR achieved absolute and relative errors of 0.54 m (7.84%) and 0.21 m³/h (13.57%), respectively, while the Feedforward Neural Network (FNN) reported errors of 0.98 m (8.24%) and 0.10 m³/h (16.71%). By contrast, the Support Vector Machine Regression (SVMR) and Generalized Additive Model (GAM) generally yielded less satisfactory prediction accuracy compared to the other methods.