To reveal the hydraulic characteristics of a centrifugal pump during variable frequency start-up，experiments were conducted on a low-specific-speed open impeller centrifugal pump under six off-design conditions. Measured parameters included inlet and outlet static pressures，flow rate，head，speed，and shaft power. The study compared the transient flow and head prediction performance of three models: Feedforward Neural Network (FFN)，Cascaded Feedforward Neural Network (CFNN)，and Multi-Layer Perceptron (MLP). Results showed that during variable frequency start-up，inlet static pressure followed a“drop-rise-drop”trend，while outlet static pressure quickly rose and stabilized. Stable values of flow rate and head decreased with increasing relative flow. The CFNN model demonstrated the best prediction accuracy.